diff --git a/litellm/proxy/management_endpoints/auto_router_endpoints.py b/litellm/proxy/management_endpoints/auto_router_endpoints.py
index 61ad998e4b3..8f89f28e07c 100644
--- a/litellm/proxy/management_endpoints/auto_router_endpoints.py
+++ b/litellm/proxy/management_endpoints/auto_router_endpoints.py
@@ -459,6 +459,10 @@ async def get_auto_router_recommendation(
         Literal["economy", "balanced", "high", "max"],
         Query(description="How close every admitted model must be to the best available quality score"),
     ] = "balanced",
+    optimize_for: Annotated[
+        Literal["cost", "task_completion_speed", "balanced"],
+        Query(description="What to minimize after the selected quality gate is applied"),
+    ] = "balanced",
     team_id: Annotated[str | None, Query(description="Team whose model access the recommendation must use")] = None,
 ) -> AutoRouterRecommendationResponse:
     """Build an editable four-tier router from only the model groups this caller can use."""
@@ -502,6 +506,7 @@ async def get_auto_router_recommendation(
             snapshot=snapshot,
             available_model_deployments=available_deployments,
             quality_level=quality_level,
+            optimize_for=optimize_for,
         )
     except ValueError as exc:
         raise HTTPException(
@@ -519,6 +524,7 @@ async def get_auto_router_recommendation(
     )
     return AutoRouterRecommendationResponse(
         quality_level=quality_level,
+        optimize_for=optimize_for,
         snapshot_id=snapshot.snapshot_id,
         snapshot_generated_at=snapshot.generated_at,
         available_model_group_count=len(available_deployments),
diff --git a/litellm/router.py b/litellm/router.py
index f33dfbba7bf..273fce66472 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -7865,6 +7865,12 @@ class Router:
             # WS session wrappers fire with result=None; per-turn costs tracked by inner calls.
             if kwargs.get("call_type") in ("_aresponses_websocket", "_arealtime"):
                 return
+            await self._record_auto_setup_response_latency(
+                kwargs=kwargs,
+                completion_response=completion_response,
+                start_time=start_time,
+                end_time=end_time,
+            )
             standard_logging_object: Final[StandardLoggingPayload | None] = kwargs.get("standard_logging_object", None)
             if standard_logging_object is None:
                 raise ValueError("standard_logging_object is None")
@@ -7976,6 +7982,56 @@ class Router:
                 "litellm.router.Router::deployment_callback_on_success(): Exception occured - %s", e
             )
 
+    async def _record_auto_setup_response_latency(
+        self,
+        *,
+        kwargs: Mapping[str, object],
+        completion_response: object,
+        start_time: object,
+        end_time: object,
+    ) -> None:
+        try:
+            metadata: Final = _SESSION_ADAPTER.validate_python(
+                get_litellm_metadata_from_kwargs(
+                    dict(kwargs)  # mutable-ok: legacy metadata helper requires a concrete dict copy
+                )
+            )
+            decision: Final = _SESSION_ADAPTER.validate_python(metadata.get("routing_decision"))
+        except ValidationError:
+            return
+        if decision.get("auto_setup_selection_mode") != "runtime_response_latency":
+            return
+        router_model_name = decision.get("router_model_name")
+        tier = decision.get("tier")
+        routed_model = decision.get("routed_model")
+        if not isinstance(router_model_name, str) or not router_model_name:
+            return
+        if not isinstance(tier, str) or not tier:
+            return
+        if not isinstance(routed_model, str) or not routed_model:
+            return
+        registered: Final = self.complexity_routers.get(router_model_name)
+        if not registered or not any(tagged.strategy.config.auto_setup is not None for tagged in registered):
+            return
+
+        from litellm.router_strategy.complexity_router.response_latency import (
+            record_runtime_response_latency,
+        )
+
+        try:
+            await record_runtime_response_latency(
+                router_cache=self.cache,
+                router_model_name=router_model_name,
+                tier=tier,
+                routed_model=routed_model,
+                kwargs=kwargs,
+                response_obj=completion_response,
+                start_time=start_time,
+                end_time=end_time,
+            )
+        except Exception as exc:  # noqa: BLE001 -- optional telemetry must never fail a successful completion
+            verbose_router_logger.warning("Could not record Auto setup response latency: %s", exc)
+
     def sync_deployment_callback_on_success(
         self,
         kwargs,  # kwargs to completion
diff --git a/litellm/router_strategy/complexity_router/README.md b/litellm/router_strategy/complexity_router/README.md
index 39078c2f304..d36ce9c2e9a 100644
--- a/litellm/router_strategy/complexity_router/README.md
+++ b/litellm/router_strategy/complexity_router/README.md
@@ -71,31 +71,43 @@ still resolve to a deployment in `model_list`; this configuration does not creat
 ### Zero-setup Auto configuration
 
 The dashboard's Add Router flow can generate the four tiers from the model groups the
-current admin or team can actually use. Auto setup asks for one product choice: Quality
-level (Economy, Balanced, High, or Max).
+current admin or team can actually use. Auto setup asks for two product choices:
+
+- Quality level: Economy, Balanced, High, or Max
+- Optimize for: Cost, Task completion speed, or Balanced
 
 The quality level is a gate, not an optimization weight. For every complexity tier,
 LiteLLM finds the best available model's conservative LiveBench quality score and admits
 only models within the selected maximum regret: 15 percentage points for Economy, 7 for
-Balanced, 3 for High, and 1 for Max. Auto then chooses the lowest estimated cost per
-completed task among the admitted models. If none of those models has comparable token
-pricing, Auto chooses the highest conservative quality score instead of failing setup.
+Balanced, 3 for High, and 1 for Max. The objective then ranks only those admitted models.
 This means Max still sends trivial work to a smaller model when that model is within one
 point of the best available quality for trivial work.
 
-Cost is estimated dollars per completed task, not input or output token price. LiteLLM
-combines the benchmark request's token mix and completion probability with the configured
-deployment's input, cache-read, and output rates. When a model group has multiple
-deployments, Auto uses the most conservative estimate.
+Cost is benchmark cost per completed task, not input or output token price. It is observed
+request dollars divided by quality-adjusted completed questions in the same complexity
+bucket, with partial benchmark credit counted as a fractional completion.
 
-Auto selects one model group for each tier and returns a normal, editable `heuristic_v2`
-configuration. It does not add a request-time policy, collect live latency observations,
-or change how the complexity router runs after creation. Existing routers, manual
-configurations, and LiteLLM's deployment-level routing are therefore unchanged.
+Task completion speed uses two forms of evidence:
 
-The bundled benchmark data is a versioned snapshot. Creating or regenerating a router uses
-the snapshot shipped with that LiteLLM version; the saved router contains only the four
-selected tiers and ordinary complexity-router settings.
+- SIMPLE and MEDIUM tiers learn equal-output response time from this proxy's successful
+  requests: time to first token plus the time to generate the expected number of visible
+  output tokens. Reasoning tokens are excluded. The policy explores every quality-admitted
+  model twice before using a rolling seven-day sample
+- COMPLEX and REASONING tiers use retry-adjusted Terminal-Bench wall-clock time to a
+  successful task. Results are compared only inside the same agent and benchmark cohort
+  and fall back to cost when comparable speed evidence is unavailable
+
+Balanced minimizes the combined log-normalized cost and completion-time scores after the
+same quality gate. The recommendation stores its snapshot ID, objective, admitted models,
+and rank order directly in the router configuration, so the generated result is visible
+and editable. Changing a generated tier's model pool in the dashboard removes that Auto
+selection policy and returns the tier to normal manual pool selection.
+
+These rules apply only when a router is created through Auto setup. Existing routers,
+manual configurations, and LiteLLM's deployment-level latency routing are unchanged. The
+bundled benchmark data is a versioned snapshot: creating or regenerating a router uses the
+snapshot shipped with that LiteLLM version, while an already saved router keeps its stored
+policy.
 
 ### Heuristic v2
 
diff --git a/litellm/router_strategy/complexity_router/artifacts/auto_router_snapshot_v0.json b/litellm/router_strategy/complexity_router/artifacts/auto_router_snapshot_v0.json
index 7be71f4b34d..67f478713b5 100644
--- a/litellm/router_strategy/complexity_router/artifacts/auto_router_snapshot_v0.json
+++ b/litellm/router_strategy/complexity_router/artifacts/auto_router_snapshot_v0.json
@@ -1,8 +1,12 @@
 {
   "activation_scope": {
+    "latency_objectives": [
+      "task_completion_speed",
+      "balanced"
+    ],
     "setup_mode": "auto"
   },
-  "artifact_sha256": "e191b5f7225e4da7a178eb5fe8780bcea24fb0166be23e2069338c6fe5522cd5",
+  "artifact_sha256": "bab7488212d40d96c5e43f80893c5ca2961d0e25160cb1aebbef36a2422b8a2b",
   "complexity_policy": {
     "cumulative_shares": {
       "complex": 1.0,
@@ -10,21 +14,36 @@
       "standard": 0.75,
       "trivial": 0.25
     },
-    "method": "relative_quantiles"
+    "method": "relative_quantiles",
+    "online_response_latency_formula": "time_to_first_token_ms + expected_visible_output_tokens / output_tokens_per_ms",
+    "speed_metric_by_complexity": {
+      "complex": "offline_task_completion_latency",
+      "simple": "online_response_latency",
+      "standard": "offline_task_completion_latency",
+      "trivial": "online_response_latency"
+    }
   },
   "definitions": {
     "cost": "Provider-adjustable estimate of dollars per completed task using the benchmark trajectory's observed token mix and completion rate.",
+    "online_response_latency": "Time to first token plus expected visible output generation time from LiteLLM's live observations.",
     "quality": "Mean LiveBench subtask score within the inferred complexity bucket.",
-    "quality_lower_bound": "Mean subtask score minus 1.96 standard errors, floored at zero."
+    "quality_lower_bound": "Mean subtask score minus 1.96 standard errors, floored at zero.",
+    "task_completion_speed": "Terminal-Bench expected wall-clock time to a successful task, including failed attempts."
   },
   "generated_at": "2026-09-03T20:00:00Z",
   "limitations": [
-    "The snapshot is an offline setup prior; it does not change runtime routing, endpoint health checks, or deployment selection.",
+    "The snapshot is an offline prior; it does not replace live endpoint health or deployment latency routing.",
+    "Terminal-Bench duration is full agent-task wall time and is only compared inside the same agent/version cohort.",
+    "Trivial and simple requests require online response-latency evidence; long-horizon benchmark time is never used to rank them.",
+    "A reference speed route anchors on the cohort containing the highest-quality measured candidate; it never compares raw seconds across cohorts.",
+    "When no quality-admitted model has comparable speed evidence, speed and balanced routes explicitly fall back to cost.",
     "LiveBench cost per completion treats partial benchmark credit as a fractional completion.",
     "LiveBench input-token means are model-level, not complexity-bucket-specific; output-token means are weighted inside each bucket.",
-    "Provider-adjusted completion cost scales the observed request cost by the benchmark token mix and source prices; cache discounts and non-token charges can make it approximate.",
-    "LiteLLM recomputes every quality gate after intersecting the snapshot with the user's configured, routable models.",
-    "Benchmark variants include their measured reasoning effort; setup preserves required effort parameters when the provider supports them."
+    "Provider-adjusted LiveBench cost scales the observed request cost by the benchmark token mix and source prices; cache discounts and non-token charges can make it approximate.",
+    "Provider-adjusted Terminal-Bench cost uses uncached input, cache-read input, and output token means; non-token tool charges are not represented.",
+    "Terminal-Bench token totals span many calls, so per-call long-context thresholds cannot be reconstructed and are not applied to trajectory totals.",
+    "Reference routes use the full benchmark catalog; LiteLLM must recompute them after intersecting with the user's configured and capability-compatible models.",
+    "Benchmark variants include their measured reasoning effort; runtime mapping must not silently claim the score for a different effort setting."
   ],
   "models": {
     "claude-fable-5-1-max-effort": {
@@ -112,6 +131,96 @@
           "source_output_cost_per_token": 5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-fable-5-1-max-effort",
+            "cohort_id": "claude-code@2.1.257",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 170.09694079545451,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 22.012545279411764,
+            "mean_cache_read_input_tokens": 5441945.505882353,
+            "mean_output_tokens": 161621.48235294118,
+            "mean_uncached_input_tokens": 1034091.7176470588,
+            "observed_duration_p50_ms": 2313576.159,
+            "observed_duration_p95_ms": 19944972.068,
+            "quality_lower_bound": 0.07382276936875576,
+            "retry_adjusted_completion_ms": 36630585.75418182,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.12941176470588237,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-fable-5-1"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-fable-5-1-max-effort",
+            "cohort_id": "claude-code@2.1.257",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 25.25749088095238,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 18.720257947058823,
+            "mean_cache_read_input_tokens": 7568509.458823529,
+            "mean_output_tokens": 220892.04705882352,
+            "mean_uncached_input_tokens": 538071.4941176471,
+            "observed_duration_p50_ms": 2841125.8109999998,
+            "observed_duration_p95_ms": 9005283.476,
+            "quality_lower_bound": 0.6390752485008677,
+            "retry_adjusted_completion_ms": 5309329.376857143,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.7411764705882353,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-fable-5-1"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-fable-5-1-max-effort",
+            "cohort_id": "claude-code@2.1.257",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 27.663567212962963,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 18.672907868750002,
+            "mean_cache_read_input_tokens": 8005376.575,
+            "mean_output_tokens": 218323.6875,
+            "mean_uncached_input_tokens": 483923.825,
+            "observed_duration_p50_ms": 2673098.842,
+            "observed_duration_p95_ms": 7985724.741,
+            "quality_lower_bound": 0.5664032172496735,
+            "retry_adjusted_completion_ms": 4729345.715351852,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.675,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-fable-5-1"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-fable-5-1-max-effort",
+            "cohort_id": "claude-code@2.1.257",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 20.434632119047624,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 16.092272793750002,
+            "mean_cache_read_input_tokens": 9009226.35,
+            "mean_output_tokens": 164275.2875,
+            "mean_uncached_input_tokens": 546191.3375,
+            "observed_duration_p50_ms": 2008713.561,
+            "observed_duration_p95_ms": 11603729.906000001,
+            "quality_lower_bound": 0.6857763938182242,
+            "retry_adjusted_completion_ms": 4635716.655238096,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.7875,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-fable-5-1"
+          }
+        ]
       }
     },
     "claude-fable-5-max-effort": {
@@ -200,6 +309,96 @@
           "source_output_cost_per_token": 5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-fable-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 392.82553275,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 23.107384279411765,
+            "mean_cache_read_input_tokens": 8784759.376470588,
+            "mean_output_tokens": 149270.90588235293,
+            "mean_uncached_input_tokens": 777170.2588235295,
+            "observed_duration_p50_ms": 2042428.4970000002,
+            "observed_duration_p95_ms": 17828109.054,
+            "quality_lower_bound": 0.025384785707111474,
+            "retry_adjusted_completion_ms": 83197786.8262,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.058823529411764705,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-fable-5"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-fable-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 33.590369561320756,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 20.94458337352941,
+            "mean_cache_read_input_tokens": 11229777.952941176,
+            "mean_output_tokens": 204881.6,
+            "mean_uncached_input_tokens": 336043.8823529412,
+            "observed_duration_p50_ms": 3535340.2729999996,
+            "observed_duration_p95_ms": 9474656.393,
+            "quality_lower_bound": 0.5172971118468915,
+            "retry_adjusted_completion_ms": 6346110.95354717,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.6235294117647059,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-fable-5"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-fable-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 58.93160589516129,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 22.835997284375,
+            "mean_cache_read_input_tokens": 11889124.975,
+            "mean_output_tokens": 173329.6125,
+            "mean_uncached_input_tokens": 402044.525,
+            "observed_duration_p50_ms": 2997084.142,
+            "observed_duration_p95_ms": 10159133.293000001,
+            "quality_lower_bound": 0.2882439968827771,
+            "retry_adjusted_completion_ms": 9369974.166258065,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.3875,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-fable-5"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-fable-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 29.201946306034483,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 21.171411071875,
+            "mean_cache_read_input_tokens": 11396242.775,
+            "mean_output_tokens": 183132.1625,
+            "mean_uncached_input_tokens": 446857.5875,
+            "observed_duration_p50_ms": 3460185.4905000003,
+            "observed_duration_p95_ms": 10187401.28,
+            "quality_lower_bound": 0.618557347158301,
+            "retry_adjusted_completion_ms": 5933162.700517242,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.725,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-fable-5"
+          }
+        ]
       }
     },
     "claude-opus-4-5-20251101-thinking-64k-high-effort": {
@@ -281,6 +480,12 @@
           "source_output_cost_per_token": 2.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "claude-opus-4-6-thinking-auto-high-effort": {
@@ -365,6 +570,12 @@
           "source_output_cost_per_token": 2.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "claude-opus-4-7-xhigh-effort": {
@@ -456,6 +667,12 @@
           "source_output_cost_per_token": 2.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "claude-opus-4-8-max-effort": {
@@ -549,6 +766,96 @@
           "source_output_cost_per_token": 2.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-opus-4-8-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 658.821080875,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 15.501672491176471,
+            "mean_cache_read_input_tokens": 14129125.54117647,
+            "mean_output_tokens": 198919.23529411765,
+            "mean_uncached_input_tokens": 554347.0235294118,
+            "observed_duration_p50_ms": 2628999.225,
+            "observed_duration_p95_ms": 17801145.046,
+            "quality_lower_bound": 0.006476375454219535,
+            "retry_adjusted_completion_ms": 210516238.57050002,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.023529411764705882,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-opus-4-8"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-opus-4-8-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 80.75515165909091,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 20.901333370588237,
+            "mean_cache_read_input_tokens": 20194287.02352941,
+            "mean_output_tokens": 302960.2117647059,
+            "mean_uncached_input_tokens": 516850.5411764706,
+            "observed_duration_p50_ms": 4935654.733,
+            "observed_duration_p95_ms": 10689830.749,
+            "quality_lower_bound": 0.1775797328865609,
+            "retry_adjusted_completion_ms": 20200870.892045453,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.25882352941176473,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-opus-4-8"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-opus-4-8-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 193.26893900000002,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 16.9110321625,
+            "mean_cache_read_input_tokens": 14685576.6375,
+            "mean_output_tokens": 236491.775,
+            "mean_uncached_input_tokens": 585015.2125,
+            "observed_duration_p50_ms": 3971969.2490000003,
+            "observed_duration_p95_ms": 8566031.918000001,
+            "quality_lower_bound": 0.043031454647439876,
+            "retry_adjusted_completion_ms": 48481955.61957143,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.0875,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-opus-4-8"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-opus-4-8-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 43.279292521276595,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 25.42658435625,
+            "mean_cache_read_input_tokens": 25549236.875,
+            "mean_output_tokens": 340427.675,
+            "mean_uncached_input_tokens": 661020.725,
+            "observed_duration_p50_ms": 6080651.175,
+            "observed_duration_p95_ms": 12414037.945,
+            "quality_lower_bound": 0.47803841126883023,
+            "retry_adjusted_completion_ms": 10616437.123489361,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.5875,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-opus-4-8"
+          }
+        ]
       }
     },
     "claude-opus-5-max-effort": {
@@ -640,6 +947,96 @@
           "source_output_cost_per_token": 2.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-opus-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 281.3378472,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 16.549285129411764,
+            "mean_cache_read_input_tokens": 15744333.729411764,
+            "mean_output_tokens": 173575.88235294117,
+            "mean_uncached_input_tokens": 694124.9764705882,
+            "observed_duration_p50_ms": 2207406.744,
+            "observed_duration_p95_ms": 17664645.654,
+            "quality_lower_bound": 0.025384785707111474,
+            "retry_adjusted_completion_ms": 94611233.12079999,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.058823529411764705,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-opus-5"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "claude-opus-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 30.785845800847458,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 21.36899885,
+            "mean_cache_read_input_tokens": 22501503.37647059,
+            "mean_output_tokens": 253603.1294117647,
+            "mean_uncached_input_tokens": 604559.9529411765,
+            "observed_duration_p50_ms": 3380587.579,
+            "observed_duration_p95_ms": 15045314.204,
+            "quality_lower_bound": 0.5895400497870746,
+            "retry_adjusted_completion_ms": 7657584.832338983,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.6941176470588235,
+            "task_count": 17,
+            "terminal_bench_model_id": "anthropic/claude-opus-5"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-opus-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 35.05708062195122,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 17.96675381875,
+            "mean_cache_read_input_tokens": 20685562.3125,
+            "mean_output_tokens": 191914.2625,
+            "mean_uncached_input_tokens": 451568.2,
+            "observed_duration_p50_ms": 3494617.0105,
+            "observed_duration_p95_ms": 7178773.611,
+            "quality_lower_bound": 0.4049314460370689,
+            "retry_adjusted_completion_ms": 7639667.404512195,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.5125,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-opus-5"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "claude-opus-5-max-effort",
+            "cohort_id": "claude-code@2.1.231",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 20.582428286764706,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 17.49506404375,
+            "mean_cache_read_input_tokens": 17807508.9,
+            "mean_output_tokens": 185969.275,
+            "mean_uncached_input_tokens": 630809.3875,
+            "observed_duration_p50_ms": 2546369.6345,
+            "observed_duration_p95_ms": 10659490.976,
+            "quality_lower_bound": 0.7558657249889396,
+            "retry_adjusted_completion_ms": 5383495.158191176,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.85,
+            "task_count": 16,
+            "terminal_bench_model_id": "anthropic/claude-opus-5"
+          }
+        ]
       }
     },
     "claude-sonnet-4-6-thinking-auto-medium-effort": {
@@ -731,6 +1128,12 @@
           "source_output_cost_per_token": 1.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "claude-sonnet-5-xhigh-effort": {
@@ -824,6 +1227,12 @@
           "source_output_cost_per_token": 1.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "deepseek-v4-flash": {
@@ -917,6 +1326,12 @@
           "source_output_cost_per_token": 2.8e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "deepseek-v4-flash-0731": {
@@ -1007,6 +1422,12 @@
           "source_output_cost_per_token": 2.8e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "deepseek-v4-flash-vision-exp": {
@@ -1088,6 +1509,12 @@
           "source_output_cost_per_token": 6.6e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "deepseek-v4-pro": {
@@ -1180,6 +1607,12 @@
           "source_output_cost_per_token": 8.7e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "deepseek-v4-pro-0813": {
@@ -1263,6 +1696,12 @@
           "source_output_cost_per_token": 8.7e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gemini-3.1-pro-preview-high": {
@@ -1347,6 +1786,12 @@
           "source_output_cost_per_token": 1.2e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gemini-3.5-flash-high": {
@@ -1431,6 +1876,12 @@
           "source_output_cost_per_token": 9e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gemini-3.5-flash-lite-high": {
@@ -1514,6 +1965,12 @@
           "source_output_cost_per_token": 2.5e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gemini-3.6-flash-high": {
@@ -1597,6 +2054,12 @@
           "source_output_cost_per_token": 7.5e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gemini-3.7-flash-high": {
@@ -1681,6 +2144,96 @@
           "source_output_cost_per_token": 3.75e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gemini-3.7-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 295.9275225750001,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 3.4815002655882368,
+            "mean_cache_read_input_tokens": 28612753.65882353,
+            "mean_output_tokens": 116238.91764705883,
+            "mean_uncached_input_tokens": 1199530.4,
+            "observed_duration_p50_ms": 648803.339,
+            "observed_duration_p95_ms": 6388953.142,
+            "quality_lower_bound": 0.0020797335877093272,
+            "retry_adjusted_completion_ms": 167815171.641,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.011764705882352941,
+            "task_count": 17,
+            "terminal_bench_model_id": "gemini/gemini-3.7-flash"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gemini-3.7-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 51.97473815625,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 2.445870030882353,
+            "mean_cache_read_input_tokens": 17370307.94117647,
+            "mean_output_tokens": 154050.90588235293,
+            "mean_uncached_input_tokens": 753874.7176470588,
+            "observed_duration_p50_ms": 1034465.7279999999,
+            "observed_duration_p95_ms": 2791459.0560000003,
+            "quality_lower_bound": 0.018449962264952782,
+            "retry_adjusted_completion_ms": 25638198.509499997,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.047058823529411764,
+            "task_count": 17,
+            "terminal_bench_model_id": "gemini/gemini-3.7-flash"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gemini-3.7-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 87.11567227500001,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 4.355783613750001,
+            "mean_cache_read_input_tokens": 39359057.1,
+            "mean_output_tokens": 159510.375,
+            "mean_uncached_input_tokens": 1074253.9,
+            "observed_duration_p50_ms": 1048099.63,
+            "observed_duration_p95_ms": 6347074.634,
+            "quality_lower_bound": 0.019613237127295572,
+            "retry_adjusted_completion_ms": 33926894.975499995,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.05,
+            "task_count": 16,
+            "terminal_bench_model_id": "gemini/gemini-3.7-flash"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gemini-3.7-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 14.627877795535715,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 5.1197572284375,
+            "mean_cache_read_input_tokens": 45995506.8375,
+            "mean_output_tokens": 204981.7,
+            "mean_uncached_input_tokens": 1201883.7875,
+            "observed_duration_p50_ms": 1198424.5915,
+            "observed_duration_p95_ms": 6527013.723,
+            "quality_lower_bound": 0.2545439440982969,
+            "retry_adjusted_completion_ms": 5402727.119178573,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.35,
+            "task_count": 16,
+            "terminal_bench_model_id": "gemini/gemini-3.7-flash"
+          }
+        ]
       }
     },
     "gemini-3.8-flash-high": {
@@ -1764,6 +2317,96 @@
           "source_output_cost_per_token": 3.75e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gemini-3.8-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 0.0,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.0,
+            "mean_cache_read_input_tokens": 47160828.941176474,
+            "mean_output_tokens": 178142.89411764705,
+            "mean_uncached_input_tokens": 1274647.8823529412,
+            "observed_duration_p50_ms": 989905.341,
+            "observed_duration_p95_ms": 14753648.89,
+            "quality_lower_bound": 0.025384785707111474,
+            "retry_adjusted_completion_ms": 41792359.309,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.058823529411764705,
+            "task_count": 17,
+            "terminal_bench_model_id": "gemini/gemini-3.8-flash"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gemini-3.8-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 0.0,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.0,
+            "mean_cache_read_input_tokens": 28410489.37647059,
+            "mean_output_tokens": 182701.0823529412,
+            "mean_uncached_input_tokens": 972770.9411764706,
+            "observed_duration_p50_ms": 1238644.2179999999,
+            "observed_duration_p95_ms": 2836649.066,
+            "quality_lower_bound": 0.1193204211030341,
+            "retry_adjusted_completion_ms": 7225142.85575,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.18823529411764706,
+            "task_count": 17,
+            "terminal_bench_model_id": "gemini/gemini-3.8-flash"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gemini-3.8-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 0.0,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.0,
+            "mean_cache_read_input_tokens": 46424741.4,
+            "mean_output_tokens": 203380.9375,
+            "mean_uncached_input_tokens": 1252358.625,
+            "observed_duration_p50_ms": 1398407.2505,
+            "observed_duration_p95_ms": 5482899.748,
+            "quality_lower_bound": 0.02698897769798554,
+            "retry_adjusted_completion_ms": 28856834.1502,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.0625,
+            "task_count": 16,
+            "terminal_bench_model_id": "gemini/gemini-3.8-flash"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gemini-3.8-flash-high",
+            "cohort_id": "mini-swe-agent@2.4.6",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 0.0,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.0,
+            "mean_cache_read_input_tokens": 81941825.4875,
+            "mean_output_tokens": 263928.3875,
+            "mean_uncached_input_tokens": 1719578.9125,
+            "observed_duration_p50_ms": 1614114.02,
+            "observed_duration_p95_ms": 5601424.65,
+            "quality_lower_bound": 0.3574781238878701,
+            "retry_adjusted_completion_ms": 5122828.089918919,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.4625,
+            "task_count": 16,
+            "terminal_bench_model_id": "gemini/gemini-3.8-flash"
+          }
+        ]
       }
     },
     "glm-5.2": {
@@ -1854,6 +2497,12 @@
           "source_output_cost_per_token": 4.4e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "glm-5.3": {
@@ -1936,6 +2585,12 @@
           "source_output_cost_per_token": 4.4e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "glm-5.3-flash": {
@@ -2017,6 +2672,12 @@
           "source_output_cost_per_token": 5e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.2-2025-12-11-high": {
@@ -2099,6 +2760,12 @@
           "source_output_cost_per_token": 1.4e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.2-codex": {
@@ -2181,6 +2848,12 @@
           "source_output_cost_per_token": 1.4e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.4-mini-xhigh": {
@@ -2264,6 +2937,12 @@
           "source_output_cost_per_token": 4.5e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.4-nano-xhigh": {
@@ -2347,6 +3026,12 @@
           "source_output_cost_per_token": 1.25e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.4-xhigh": {
@@ -2436,6 +3121,12 @@
           "source_output_cost_per_token": 1.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.5-xhigh": {
@@ -2522,6 +3213,12 @@
           "source_output_cost_per_token": 3e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "gpt-5.6-luna-max": {
@@ -2610,6 +3307,96 @@
           "source_output_cost_per_token": 6e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gpt-5.6-luna-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": null,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.6297226068235294,
+            "mean_cache_read_input_tokens": 21967651.752941176,
+            "mean_output_tokens": 95748.37647058823,
+            "mean_uncached_input_tokens": 309517.24705882353,
+            "observed_duration_p50_ms": 812183.1919999999,
+            "observed_duration_p95_ms": 30386749.601,
+            "quality_lower_bound": 0.0,
+            "retry_adjusted_completion_ms": null,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.0,
+            "task_count": 17,
+            "terminal_bench_model_id": "openai/gpt-5.6-luna"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gpt-5.6-luna-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 3.6673198505882363,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.7334639701176473,
+            "mean_cache_read_input_tokens": 21637394.447058823,
+            "mean_output_tokens": 168592.02352941176,
+            "mean_uncached_input_tokens": 397313.6705882353,
+            "observed_duration_p50_ms": 1451894.867,
+            "observed_duration_p95_ms": 9231407.450000001,
+            "quality_lower_bound": 0.12878892241565565,
+            "retry_adjusted_completion_ms": 14908957.570411764,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.2,
+            "task_count": 17,
+            "terminal_bench_model_id": "openai/gpt-5.6-luna"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gpt-5.6-luna-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 16.9274472,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 0.63477927,
+            "mean_cache_read_input_tokens": 18678964.6875,
+            "mean_output_tokens": 130148.925,
+            "mean_uncached_input_tokens": 422307.325,
+            "observed_duration_p50_ms": 1577440.2845,
+            "observed_duration_p95_ms": 5629013.643,
+            "quality_lower_bound": 0.012834336300788755,
+            "retry_adjusted_completion_ms": 54544322.523,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.0375,
+            "task_count": 16,
+            "terminal_bench_model_id": "openai/gpt-5.6-luna"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gpt-5.6-luna-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 4.926968522203215,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 2.2787229415189874,
+            "mean_cache_read_input_tokens": null,
+            "mean_output_tokens": null,
+            "mean_uncached_input_tokens": null,
+            "observed_duration_p50_ms": 3137410.6835000003,
+            "observed_duration_p95_ms": 28950832.13,
+            "quality_lower_bound": 0.3574781238878701,
+            "retry_adjusted_completion_ms": 16030999.12881081,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.4625,
+            "task_count": 16,
+            "terminal_bench_model_id": "openai/gpt-5.6-luna"
+          }
+        ]
       }
     },
     "gpt-5.6-sol-max": {
@@ -2697,6 +3484,96 @@
           "source_output_cost_per_token": 3e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gpt-5.6-sol-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 254.19713079999997,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 8.971663439999999,
+            "mean_cache_read_input_tokens": 16735234.835294118,
+            "mean_output_tokens": 62267.282352941176,
+            "mean_uncached_input_tokens": 209666.14117647058,
+            "observed_duration_p50_ms": 1013924.23,
+            "observed_duration_p95_ms": 18501705.555,
+            "quality_lower_bound": 0.012074823021218326,
+            "retry_adjusted_completion_ms": 105534731.96500002,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.03529411764705882,
+            "task_count": 17,
+            "terminal_bench_model_id": "openai/gpt-5.6-sol"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gpt-5.6-sol-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 14.184279841379311,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 4.839342534117647,
+            "mean_cache_read_input_tokens": 7048592.423529412,
+            "mean_output_tokens": 66036.84705882354,
+            "mean_uncached_input_tokens": 140110.48235294118,
+            "observed_duration_p50_ms": 1373627.158,
+            "observed_duration_p95_ms": 3545131.2939999998,
+            "quality_lower_bound": 0.2492177176176207,
+            "retry_adjusted_completion_ms": 4732196.68475862,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.3411764705882353,
+            "task_count": 17,
+            "terminal_bench_model_id": "openai/gpt-5.6-sol"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gpt-5.6-sol-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 21.92680297777778,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 7.400296005000001,
+            "mean_cache_read_input_tokens": 12077936.6375,
+            "mean_output_tokens": 69831.1875,
+            "mean_uncached_input_tokens": 236228.9,
+            "observed_duration_p50_ms": 1280606.867,
+            "observed_duration_p95_ms": 7725407.675000001,
+            "quality_lower_bound": 0.24345459812221978,
+            "retry_adjusted_completion_ms": 5344111.560555556,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.3375,
+            "task_count": 16,
+            "terminal_bench_model_id": "openai/gpt-5.6-sol"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gpt-5.6-sol-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 13.002322996874996,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 10.401858397499998,
+            "mean_cache_read_input_tokens": 18088253.65,
+            "mean_output_tokens": 89979.0625,
+            "mean_uncached_input_tokens": 278267.675,
+            "observed_duration_p50_ms": 1556719.1935,
+            "observed_duration_p95_ms": 8823486.585,
+            "quality_lower_bound": 0.6995355220341885,
+            "retry_adjusted_completion_ms": 3411662.1276875,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.8,
+            "task_count": 16,
+            "terminal_bench_model_id": "openai/gpt-5.6-sol"
+          }
+        ]
       }
     },
     "gpt-5.6-terra-max": {
@@ -2785,6 +3662,96 @@
           "source_output_cost_per_token": 1.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gpt-5.6-terra-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 207.23259980000006,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 4.876061171764707,
+            "mean_cache_read_input_tokens": 16777721.388235293,
+            "mean_output_tokens": 79003.21176470588,
+            "mean_uncached_input_tokens": 231868.31764705884,
+            "observed_duration_p50_ms": 760895.202,
+            "observed_duration_p95_ms": 17511504.763,
+            "quality_lower_bound": 0.006476375454219535,
+            "retry_adjusted_completion_ms": 143084489.17000002,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.023529411764705882,
+            "task_count": 17,
+            "terminal_bench_model_id": "openai/gpt-5.6-terra"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "gpt-5.6-terra-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 17.41710915542328,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 3.6883289976190476,
+            "mean_cache_read_input_tokens": null,
+            "mean_output_tokens": null,
+            "mean_uncached_input_tokens": null,
+            "observed_duration_p50_ms": 1243788.852,
+            "observed_duration_p95_ms": 3643298.026,
+            "quality_lower_bound": 0.13836125681504122,
+            "retry_adjusted_completion_ms": 7933672.004000001,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.21176470588235294,
+            "task_count": 17,
+            "terminal_bench_model_id": "openai/gpt-5.6-terra"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gpt-5.6-terra-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 56.02047092,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 3.5012794325,
+            "mean_cache_read_input_tokens": 10359244.35,
+            "mean_output_tokens": 73774.925,
+            "mean_uncached_input_tokens": 218093.775,
+            "observed_duration_p50_ms": 1329513.8054999998,
+            "observed_duration_p95_ms": 3397796.914,
+            "quality_lower_bound": 0.02698897769798554,
+            "retry_adjusted_completion_ms": 23351506.7788,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.0625,
+            "task_count": 16,
+            "terminal_bench_model_id": "openai/gpt-5.6-terra"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "gpt-5.6-terra-max",
+            "cohort_id": "codex@0.149.1",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 15.850715830434785,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 9.114161602500001,
+            "mean_cache_read_input_tokens": 30301997.95,
+            "mean_output_tokens": 159034.0875,
+            "mean_uncached_input_tokens": 460952.0875,
+            "observed_duration_p50_ms": 1951635.4270000001,
+            "observed_duration_p95_ms": 14613717.367999999,
+            "quality_lower_bound": 0.465690919708387,
+            "retry_adjusted_completion_ms": 6175673.676978262,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.575,
+            "task_count": 16,
+            "terminal_bench_model_id": "openai/gpt-5.6-terra"
+          }
+        ]
       }
     },
     "grok-4.3": {
@@ -2867,6 +3834,12 @@
           "source_output_cost_per_token": 2.5e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "grok-4.5": {
@@ -2946,6 +3919,96 @@
           "source_output_cost_per_token": 6e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "grok-4.5",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": null,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 2.71251968,
+            "mean_cache_read_input_tokens": 3847177.035294118,
+            "mean_output_tokens": 52782.964705882354,
+            "mean_uncached_input_tokens": 355004.8941176471,
+            "observed_duration_p50_ms": 809437.804,
+            "observed_duration_p95_ms": 17566938.117999997,
+            "quality_lower_bound": 0.0,
+            "retry_adjusted_completion_ms": null,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.0,
+            "task_count": 17,
+            "terminal_bench_model_id": "xai/grok-4.5"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "grok-4.5",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 46.48075626666667,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 5.4683242666666665,
+            "mean_cache_read_input_tokens": null,
+            "mean_output_tokens": null,
+            "mean_uncached_input_tokens": null,
+            "observed_duration_p50_ms": 1772457.824,
+            "observed_duration_p95_ms": 5691062.943,
+            "quality_lower_bound": 0.06517287423683975,
+            "retry_adjusted_completion_ms": 22212528.7113,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.11764705882352941,
+            "task_count": 17,
+            "terminal_bench_model_id": "xai/grok-4.5"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "grok-4.5",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 130.93046981818182,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 3.2732617454545454,
+            "mean_cache_read_input_tokens": 5532003.2,
+            "mean_output_tokens": 68074.2875,
+            "mean_uncached_input_tokens": 224299.1,
+            "observed_duration_p50_ms": 1359035.736,
+            "observed_duration_p95_ms": 3994916.012,
+            "quality_lower_bound": 0.0068827488981252315,
+            "retry_adjusted_completion_ms": 62711641.6255,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.025,
+            "task_count": 16,
+            "terminal_bench_model_id": "xai/grok-4.5"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "grok-4.5",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 43.48064595862069,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 15.76173416,
+            "mean_cache_read_input_tokens": null,
+            "mean_output_tokens": null,
+            "mean_uncached_input_tokens": null,
+            "observed_duration_p50_ms": 2102327.9115000004,
+            "observed_duration_p95_ms": 28832497.533,
+            "quality_lower_bound": 0.26570633724259446,
+            "retry_adjusted_completion_ms": 14499273.850896552,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.3625,
+            "task_count": 16,
+            "terminal_bench_model_id": "xai/grok-4.5"
+          }
+        ]
       }
     },
     "grok-4.6": {
@@ -3028,6 +4091,96 @@
           "source_output_cost_per_token": 6e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "grok-4.6",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "complex",
+            "cost_per_completed_task_usd": 80.149236,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 3.7717287529411765,
+            "mean_cache_read_input_tokens": 4618074.352941177,
+            "mean_output_tokens": 71291.23529411765,
+            "mean_uncached_input_tokens": 195723.57647058825,
+            "observed_duration_p50_ms": 1115803.779,
+            "observed_duration_p95_ms": 15627159.765,
+            "quality_lower_bound": 0.018449962264952782,
+            "retry_adjusted_completion_ms": 58126086.48225,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.047058823529411764,
+            "task_count": 17,
+            "terminal_bench_model_id": "xai/grok-4.6"
+          }
+        ],
+        "simple": [
+          {
+            "attempt_count": 85,
+            "benchmark_model_id": "grok-4.6",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "simple",
+            "cost_per_completed_task_usd": 39.875287432228916,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 7.505936457831325,
+            "mean_cache_read_input_tokens": 8020312.847058823,
+            "mean_output_tokens": 109234.10588235295,
+            "mean_uncached_input_tokens": 243557.15294117646,
+            "observed_duration_p50_ms": 1821589.598,
+            "observed_duration_p95_ms": 4264177.431,
+            "quality_lower_bound": 0.1193204211030341,
+            "retry_adjusted_completion_ms": 10645207.7405625,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.18823529411764706,
+            "task_count": 17,
+            "terminal_bench_model_id": "xai/grok-4.6"
+          }
+        ],
+        "standard": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "grok-4.6",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "standard",
+            "cost_per_completed_task_usd": 472.3057235443037,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 11.807643088607593,
+            "mean_cache_read_input_tokens": 12985204.8,
+            "mean_output_tokens": 93418.95,
+            "mean_uncached_input_tokens": 287172.2375,
+            "observed_duration_p50_ms": 1404832.459,
+            "observed_duration_p95_ms": 7392303.36,
+            "quality_lower_bound": 0.0068827488981252315,
+            "retry_adjusted_completion_ms": 74351417.9385,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.025,
+            "task_count": 16,
+            "terminal_bench_model_id": "xai/grok-4.6"
+          }
+        ],
+        "trivial": [
+          {
+            "attempt_count": 80,
+            "benchmark_model_id": "grok-4.6",
+            "cohort_id": "grok-build@1.0.5",
+            "complexity": "trivial",
+            "cost_per_completed_task_usd": 38.115191288888894,
+            "excluded_infrastructure_trials": 0,
+            "mean_attempt_cost_usd": 21.4397951,
+            "mean_cache_read_input_tokens": 22056990.4,
+            "mean_output_tokens": 122500.6375,
+            "mean_uncached_input_tokens": 391925.475,
+            "observed_duration_p50_ms": 1962060.6230000001,
+            "observed_duration_p95_ms": 9825563.117,
+            "quality_lower_bound": 0.45340937868656955,
+            "retry_adjusted_completion_ms": 4988289.024244444,
+            "source_id": "terminal-bench-4.0.0",
+            "success_probability": 0.5625,
+            "task_count": 16,
+            "terminal_bench_model_id": "xai/grok-4.6"
+          }
+        ]
       }
     },
     "grok-build-0.1": {
@@ -3107,6 +4260,12 @@
           "source_output_cost_per_token": 2e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "inkling-xhigh": {
@@ -3190,6 +4349,12 @@
           "source_output_cost_per_token": 4.68e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "kimi-k2.6-thinking": {
@@ -3267,6 +4432,12 @@
           "source_output_cost_per_token": 4e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "kimi-k2.7-code": {
@@ -3356,6 +4527,12 @@
           "source_output_cost_per_token": 4e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "kimi-k3": {
@@ -3441,6 +4618,12 @@
           "source_output_cost_per_token": 1.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "minimax-m3": {
@@ -3527,6 +4710,12 @@
           "source_output_cost_per_token": 1.2e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "muse-spark-1.1-xhigh": {
@@ -3608,6 +4797,12 @@
           "source_output_cost_per_token": 4.25e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "muse-spark-1.2-xhigh": {
@@ -3689,6 +4884,12 @@
           "source_output_cost_per_token": 4.25e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "muse-spark-1.3-xhigh": {
@@ -3770,6 +4971,12 @@
           "source_output_cost_per_token": 4.25e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "nemotron-3-ultra-550b-a55b": {
@@ -3849,6 +5056,12 @@
           "source_output_cost_per_token": 2.4e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "ox-alpha-max": {
@@ -3926,6 +5139,12 @@
           "source_output_cost_per_token": 0.0,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "qwen3.6-27b": {
@@ -4009,6 +5228,12 @@
           "source_output_cost_per_token": 3.6000000000000003e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "qwen3.6-plus": {
@@ -4089,6 +5314,12 @@
           "source_output_cost_per_token": 3e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "qwen3.7-max": {
@@ -4173,6 +5404,12 @@
           "source_output_cost_per_token": 7.5e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "qwen3.8-27b": {
@@ -4254,6 +5491,12 @@
           "source_output_cost_per_token": 6.5e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "qwen3.8-flash-next": {
@@ -4331,6 +5574,12 @@
           "source_output_cost_per_token": 4.6999999999999995e-07,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "qwen3.8-max": {
@@ -4415,6 +5664,12 @@
           "source_output_cost_per_token": 6e-06,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     },
     "smaug-agentic": {
@@ -4492,6 +5747,12 @@
           "source_output_cost_per_token": 1.5e-05,
           "subtask_count": 5
         }
+      },
+      "task_completion_speed_by_complexity": {
+        "complex": [],
+        "simple": [],
+        "standard": [],
+        "trivial": []
       }
     }
   },
@@ -4520,12 +5781,145 @@
         "path": "livebench/table_2026_06_25.csv",
         "sha256": "b8e6339f7fafe6df8f6ecc8b0abe3dc529b5737a6c7edd3e24ffcb36e02fd118",
         "source_id": "livebench-2026-06-25"
+      },
+      {
+        "bytes": 15970,
+        "path": "terminal-bench/submissions/2026-08-26-anthropic-claude-fable-5-max-claude-code.json",
+        "sha256": "487f60477db9c36bd5cdb3d50c80634233bfa53de3368cf33337a771f17aa07e",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15964,
+        "path": "terminal-bench/submissions/2026-08-26-anthropic-claude-opus-4-8-max-claude-code.json",
+        "sha256": "25aea019c65af71083bb684cd26c30c4c6eb3f0c91d4d7473c7090c1b5e1d289",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15958,
+        "path": "terminal-bench/submissions/2026-08-26-anthropic-claude-opus-5-max-claude-code.json",
+        "sha256": "06fbebf9bf42c430c49b020f86b9365f5ae772c6063bab4fce98f54e349ed80c",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15932,
+        "path": "terminal-bench/submissions/2026-08-26-openai-gpt-5-6-luna-max-codex.json",
+        "sha256": "59ae3614eca2b0c1f087b7a50d65e3fc40dd2ef0188c2fab40d32c5a50f823a7",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15924,
+        "path": "terminal-bench/submissions/2026-08-26-openai-gpt-5-6-sol-max-codex.json",
+        "sha256": "b41e5f5f15aeb7657f9d32b33a868aecab6dd6db2d48b7480638455664102c5b",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15932,
+        "path": "terminal-bench/submissions/2026-08-26-openai-gpt-5-6-terra-max-codex.json",
+        "sha256": "1c7ca98885e40ed4dc31d23391982848b2c1dd87b0595743fa242ad848f1b670",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15887,
+        "path": "terminal-bench/submissions/2026-08-26-xai-grok-4-5-none-grok-build.json",
+        "sha256": "fa618ecb274776479019af6d62c538197cbb563a44d8b8a45752dde51d603b93",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15886,
+        "path": "terminal-bench/submissions/2026-08-26-xai-grok-4-6-none-grok-build.json",
+        "sha256": "31526e5dff0e06680085bb5029700b0fddc607962e7c6aea9a911600c6b2df1a",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15943,
+        "path": "terminal-bench/submissions/2026-09-01-gemini-gemini-3-8-flash-high-mini-swe-agent.json",
+        "sha256": "835f2090e7d40d9e5c6c5483da49c5f3147a0304e615df49c71aafa859a62c50",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15943,
+        "path": "terminal-bench/submissions/2026-09-02-anthropic-claude-fable-5-1-max-claude-code.json",
+        "sha256": "cccdb936c1266db66ce922db48144ed0716c1294eec3dcc1bdcf976711c92813",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 15953,
+        "path": "terminal-bench/submissions/2026-09-02-gemini-gemini-3-7-flash-high-mini-swe-agent.json",
+        "sha256": "013ace9d0c06b97aed4227b624aff0724257b2ca44a113553a03707545efc0cc",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 405512,
+        "path": "terminal-bench/trials/claude-fable-5-1-max-claude-code.json",
+        "sha256": "eace284d655bb870eeb113298b921d1ffde2e53f3c43ecbccf619cc6b2a50dd5",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 405541,
+        "path": "terminal-bench/trials/claude-fable-5-max-claude-code.json",
+        "sha256": "f6866893713a1feafe23a2136ae6334ff5cadf6887572e94ca8e33a3cfedc4da",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 406419,
+        "path": "terminal-bench/trials/claude-opus-4-8-max-claude-code.json",
+        "sha256": "feddf6a0222d4140696f0e25d72a6186282d20eb9170908692ada19e49a3a9ec",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 404974,
+        "path": "terminal-bench/trials/claude-opus-5-max-claude-code.json",
+        "sha256": "7a1bea779168226edfd73b547528a042a26d9fd573a6af7a174f90516ede2494",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 409532,
+        "path": "terminal-bench/trials/gemini-3-7-flash-high-mini-swe-agent.json",
+        "sha256": "b674fd778f9148d19cd3e38f441c4ecf168eb0f1056f5d9e75843ed80e66f6d2",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 406835,
+        "path": "terminal-bench/trials/gemini-3-8-flash-high-mini-swe-agent.json",
+        "sha256": "791ee2577ac9b9c9100616f53af6df0945d86b835bd39d7318bb7ee15ccf1f73",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 401789,
+        "path": "terminal-bench/trials/gpt-5-6-luna-max-codex.json",
+        "sha256": "9e223e1f7b86990d4260b96ccc76044a86b08df7c90b96c9e00f8a2a150134af",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 400386,
+        "path": "terminal-bench/trials/gpt-5-6-sol-max-codex.json",
+        "sha256": "68729d3d8bfe09362752a8ac802cbf7fc3bea6768494d9cab14439dbdac1bb54",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 401701,
+        "path": "terminal-bench/trials/gpt-5-6-terra-max-codex.json",
+        "sha256": "646b7879615f7ae5a9861bbbd7271416db4734e30a646fa84993e0692a714d6d",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 399870,
+        "path": "terminal-bench/trials/grok-4-5-grok-build.json",
+        "sha256": "0745d3a91ec75cd544f8c79ca6b162e74b2a19416d9ce428315604549ff65093",
+        "source_id": "terminal-bench-4.0.0"
+      },
+      {
+        "bytes": 399597,
+        "path": "terminal-bench/trials/grok-4-6-grok-build.json",
+        "sha256": "6323e4142ebb1c3a054b42340e8cffa7143ff6ddf480dfdcc89cb5a710660178",
+        "source_id": "terminal-bench-4.0.0"
       }
     ],
-    "manifest_sha256": "32def689d0fcb59afc8ca06fbc0d2f015c1268cb34dd4aaba991e1150ca437e8",
+    "manifest_sha256": "128c50d897e1d3f5749c5dd9aa4ec8fa5efb9d00dea983af8194a4981c25b557",
     "source_revisions": {
       "litellm-model-registry-2026-09-03": "828d561fcc2704eaf90617be8cdc058fccbd8508",
-      "livebench-2026-06-25": "62240f848c977d4202c1029191ac663498745f2f"
+      "livebench-2026-06-25": "62240f848c977d4202c1029191ac663498745f2f",
+      "terminal-bench-4.0.0": "83c7a6172d629c6575b785ab12c8db787bb2e323"
     }
   },
   "quality_tiers": {
@@ -4546,7 +5940,2205 @@
       "maximum_quality_regret": 0.01
     }
   },
-  "schema_version": "3.0",
+  "reference_routes": {
+    "balanced": {
+      "balanced": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "deepseek-v4-flash-vision-exp",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.5392130921684612,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "deepseek-v4-pro-0813",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-pro-0813",
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "gemini-3.6-flash-high",
+            "muse-spark-1.3-xhigh",
+            "claude-opus-5-max-effort",
+            "kimi-k3",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "glm-5.3",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-pro-0813",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "glm-5.3",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh"
+          ],
+          "quality_floor": 0.7797278750259673,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "gemini-3.1-pro-preview-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "grok-4.5",
+            "grok-4.6",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.70251553515036,
+          "selected_model_id": "claude-fable-5-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.231",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "deepseek-v4-flash-0731",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-flash",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "deepseek-v4-flash-vision-exp",
+            "gemini-3.5-flash-lite-high",
+            "kimi-k2.7-code",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-luna-max",
+            "grok-4.6",
+            "muse-spark-1.2-xhigh",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "gpt-5.6-terra-max",
+            "glm-5.2",
+            "gpt-5.2-codex",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "nemotron-3-ultra-550b-a55b",
+            "gpt-5.4-mini-xhigh",
+            "qwen3.7-max",
+            "qwen3.6-plus",
+            "inkling-xhigh",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.5-flash-lite-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-mini-xhigh",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "inkling-xhigh",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "nemotron-3-ultra-550b-a55b",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.9068668617803175,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "cost": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-vision-exp",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max",
+            "kimi-k3",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.1042448619785062,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "deepseek-v4-flash-vision-exp",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.5392130921684612,
+          "selected_model_id": "deepseek-v4-flash-vision-exp",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "deepseek-v4-pro-0813",
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "gemini-3.6-flash-high",
+            "muse-spark-1.3-xhigh",
+            "claude-opus-5-max-effort",
+            "kimi-k3",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "glm-5.3",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.03819422168482955,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-pro-0813",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "glm-5.3",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh"
+          ],
+          "quality_floor": 0.7797278750259673,
+          "selected_model_id": "deepseek-v4-pro-0813",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "grok-4.6",
+            "grok-4.5",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max",
+            "gemini-3.1-pro-preview-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.14293474305698176,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "gemini-3.1-pro-preview-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "grok-4.5",
+            "grok-4.6",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.70251553515036,
+          "selected_model_id": "grok-4.6",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-flash",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "deepseek-v4-flash-vision-exp",
+            "gemini-3.5-flash-lite-high",
+            "kimi-k2.7-code",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-luna-max",
+            "grok-4.6",
+            "muse-spark-1.2-xhigh",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "gpt-5.6-terra-max",
+            "glm-5.2",
+            "gpt-5.2-codex",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "nemotron-3-ultra-550b-a55b",
+            "gpt-5.4-mini-xhigh",
+            "qwen3.7-max",
+            "qwen3.6-plus",
+            "inkling-xhigh",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": 0.001468206744000269,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.5-flash-lite-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-mini-xhigh",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "inkling-xhigh",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "nemotron-3-ultra-550b-a55b",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.9068668617803175,
+          "selected_model_id": "deepseek-v4-flash-0731",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "task_completion_speed": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 36630585.75418182,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "deepseek-v4-flash-vision-exp",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.5392130921684612,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "deepseek-v4-pro-0813",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-pro-0813",
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "gemini-3.6-flash-high",
+            "muse-spark-1.3-xhigh",
+            "claude-opus-5-max-effort",
+            "kimi-k3",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "glm-5.3",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-pro-0813",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "glm-5.3",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh"
+          ],
+          "quality_floor": 0.7797278750259673,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 9369974.166258065,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "gemini-3.1-pro-preview-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "grok-4.5",
+            "grok-4.6",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.70251553515036,
+          "selected_model_id": "claude-fable-5-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.231",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "deepseek-v4-flash-0731",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-flash",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "deepseek-v4-flash-vision-exp",
+            "gemini-3.5-flash-lite-high",
+            "kimi-k2.7-code",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-luna-max",
+            "grok-4.6",
+            "muse-spark-1.2-xhigh",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "gpt-5.6-terra-max",
+            "glm-5.2",
+            "gpt-5.2-codex",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "nemotron-3-ultra-550b-a55b",
+            "gpt-5.4-mini-xhigh",
+            "qwen3.7-max",
+            "qwen3.6-plus",
+            "inkling-xhigh",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.5-flash-lite-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-mini-xhigh",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "inkling-xhigh",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "nemotron-3-ultra-550b-a55b",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.9068668617803175,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      }
+    },
+    "economy": {
+      "balanced": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro-0813",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.4592130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "deepseek-v4-flash-0731",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-flash-vision-exp",
+            "kimi-k2.7-code",
+            "gemini-3.7-flash-high",
+            "muse-spark-1.1-xhigh",
+            "qwen3.7-max",
+            "grok-4.6",
+            "grok-4.5",
+            "qwen3.6-plus",
+            "muse-spark-1.2-xhigh",
+            "gemini-3.6-flash-high",
+            "glm-5.2",
+            "gpt-5.2-2025-12-11-high",
+            "muse-spark-1.3-xhigh",
+            "gpt-5.6-luna-max",
+            "claude-opus-5-max-effort",
+            "gemini-3.5-flash-high",
+            "gemini-3.1-pro-preview-high",
+            "qwen3.8-max",
+            "gpt-5.2-codex",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "glm-5.3",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.6997278750259673,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-opus-5-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-8-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.7-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "grok-build-0.1",
+            "inkling-xhigh",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.6225155351503601,
+          "selected_model_id": "claude-opus-5-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.231",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "deepseek-v4-flash-0731",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-flash",
+            "minimax-m3",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "gpt-5.4-nano-xhigh",
+            "deepseek-v4-flash-vision-exp",
+            "gemini-3.5-flash-lite-high",
+            "kimi-k2.7-code",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-luna-max",
+            "grok-4.6",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "gpt-5.6-terra-max",
+            "glm-5.2",
+            "gpt-5.2-codex",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "nemotron-3-ultra-550b-a55b",
+            "gpt-5.4-mini-xhigh",
+            "qwen3.7-max",
+            "qwen3.6-plus",
+            "inkling-xhigh",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "qwen3.6-27b",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.5-flash-lite-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-mini-xhigh",
+            "gpt-5.4-nano-xhigh",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "inkling-xhigh",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "minimax-m3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "nemotron-3-ultra-550b-a55b",
+            "qwen3.6-27b",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.8268668617803175,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "cost": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "glm-5.3-flash",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-flash-vision-exp",
+            "grok-4.5",
+            "gemini-3.7-flash-high",
+            "qwen3.8-27b",
+            "muse-spark-1.3-xhigh",
+            "muse-spark-1.1-xhigh",
+            "gemini-3.8-flash-high",
+            "grok-4.6",
+            "qwen3.8-max",
+            "gpt-5.6-terra-max",
+            "kimi-k3",
+            "gpt-5.4-xhigh",
+            "glm-5.3",
+            "gpt-5.5-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.6-sol-max",
+            "muse-spark-1.2-xhigh",
+            "claude-opus-5-max-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.08444897731421191,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro-0813",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.4592130921684613,
+          "selected_model_id": "glm-5.3-flash",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-flash-vision-exp",
+            "kimi-k2.7-code",
+            "gemini-3.7-flash-high",
+            "muse-spark-1.1-xhigh",
+            "qwen3.7-max",
+            "grok-4.6",
+            "grok-4.5",
+            "qwen3.6-plus",
+            "muse-spark-1.2-xhigh",
+            "gemini-3.6-flash-high",
+            "glm-5.2",
+            "gpt-5.2-2025-12-11-high",
+            "muse-spark-1.3-xhigh",
+            "gpt-5.6-luna-max",
+            "claude-opus-5-max-effort",
+            "gemini-3.5-flash-high",
+            "gemini-3.1-pro-preview-high",
+            "qwen3.8-max",
+            "gpt-5.2-codex",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "glm-5.3",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.0060579263531826775,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.6997278750259673,
+          "selected_model_id": "deepseek-v4-flash-0731",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "grok-build-0.1",
+            "deepseek-v4-flash-0731",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-flash-vision-exp",
+            "muse-spark-1.1-xhigh",
+            "grok-4.6",
+            "grok-4.5",
+            "muse-spark-1.2-xhigh",
+            "glm-5.2",
+            "gemini-3.7-flash-high",
+            "muse-spark-1.3-xhigh",
+            "inkling-xhigh",
+            "gpt-5.2-codex",
+            "qwen3.8-max",
+            "gpt-5.2-2025-12-11-high",
+            "qwen3.7-max",
+            "gemini-3.1-pro-preview-high",
+            "gpt-5.4-xhigh",
+            "gpt-5.6-terra-max",
+            "gpt-5.5-xhigh",
+            "kimi-k3",
+            "gpt-5.6-sol-max",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "glm-5.3",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.0034918518733508376,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.7-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "grok-build-0.1",
+            "inkling-xhigh",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.6225155351503601,
+          "selected_model_id": "grok-build-0.1",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-flash",
+            "minimax-m3",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "gpt-5.4-nano-xhigh",
+            "deepseek-v4-flash-vision-exp",
+            "gemini-3.5-flash-lite-high",
+            "kimi-k2.7-code",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-luna-max",
+            "grok-4.6",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "gpt-5.6-terra-max",
+            "glm-5.2",
+            "gpt-5.2-codex",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "nemotron-3-ultra-550b-a55b",
+            "gpt-5.4-mini-xhigh",
+            "qwen3.7-max",
+            "qwen3.6-plus",
+            "inkling-xhigh",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "qwen3.6-27b",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": 0.001468206744000269,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.5-flash-lite-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-mini-xhigh",
+            "gpt-5.4-nano-xhigh",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "inkling-xhigh",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "minimax-m3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "nemotron-3-ultra-550b-a55b",
+            "qwen3.6-27b",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.8268668617803175,
+          "selected_model_id": "deepseek-v4-flash-0731",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "task_completion_speed": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 36630585.75418182,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro-0813",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.4592130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "deepseek-v4-flash-0731",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-flash-vision-exp",
+            "kimi-k2.7-code",
+            "gemini-3.7-flash-high",
+            "muse-spark-1.1-xhigh",
+            "qwen3.7-max",
+            "grok-4.6",
+            "grok-4.5",
+            "qwen3.6-plus",
+            "muse-spark-1.2-xhigh",
+            "gemini-3.6-flash-high",
+            "glm-5.2",
+            "gpt-5.2-2025-12-11-high",
+            "muse-spark-1.3-xhigh",
+            "gpt-5.6-luna-max",
+            "claude-opus-5-max-effort",
+            "gemini-3.5-flash-high",
+            "gemini-3.1-pro-preview-high",
+            "qwen3.8-max",
+            "gpt-5.2-codex",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "claude-sonnet-5-xhigh-effort",
+            "glm-5.3",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.6997278750259673,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-opus-5-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-8-max-effort"
+          ],
+          "objective_value": 7639667.404512195,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.7-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.5",
+            "grok-4.6",
+            "grok-build-0.1",
+            "inkling-xhigh",
+            "kimi-k3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.6225155351503601,
+          "selected_model_id": "claude-opus-5-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.231",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "deepseek-v4-flash-0731",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-flash-0731",
+            "glm-5.3-flash",
+            "deepseek-v4-flash",
+            "minimax-m3",
+            "qwen3.8-27b",
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "gpt-5.4-nano-xhigh",
+            "deepseek-v4-flash-vision-exp",
+            "gemini-3.5-flash-lite-high",
+            "kimi-k2.7-code",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-luna-max",
+            "grok-4.6",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "gpt-5.6-terra-max",
+            "glm-5.2",
+            "gpt-5.2-codex",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "nemotron-3-ultra-550b-a55b",
+            "gpt-5.4-mini-xhigh",
+            "qwen3.7-max",
+            "qwen3.6-plus",
+            "inkling-xhigh",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "qwen3.6-27b",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-4-6-thinking-auto-medium-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-flash",
+            "deepseek-v4-flash-0731",
+            "deepseek-v4-flash-vision-exp",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.5-flash-lite-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.2-codex",
+            "gpt-5.4-mini-xhigh",
+            "gpt-5.4-nano-xhigh",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-luna-max",
+            "gpt-5.6-sol-max",
+            "gpt-5.6-terra-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "inkling-xhigh",
+            "kimi-k2.7-code",
+            "kimi-k3",
+            "minimax-m3",
+            "muse-spark-1.1-xhigh",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "nemotron-3-ultra-550b-a55b",
+            "qwen3.6-27b",
+            "qwen3.6-plus",
+            "qwen3.7-max",
+            "qwen3.8-27b",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.8268668617803175,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      }
+    },
+    "high": {
+      "balanced": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "quality_floor": 0.5792130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "gpt-5.6-sol-max",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "gpt-5.6-sol-max",
+            "gpt-5.5-xhigh",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max"
+          ],
+          "quality_floor": 0.8197278750259673,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-max-effort",
+            "gemini-3.1-pro-preview-high",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.7425155351503601,
+          "selected_model_id": "claude-fable-5-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.231",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "deepseek-v4-pro-0813",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "qwen3.7-max",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.3",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh",
+            "qwen3.7-max",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.9468668617803175,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "cost": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 4.002891945931923,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "quality_floor": 0.5792130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "gpt-5.6-sol-max",
+            "gpt-5.5-xhigh",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.18762597508606915,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max"
+          ],
+          "quality_floor": 0.8197278750259673,
+          "selected_model_id": "gpt-5.6-sol-max",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max",
+            "gemini-3.1-pro-preview-high",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.14910282543392475,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-max-effort",
+            "gemini-3.1-pro-preview-high",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.7425155351503601,
+          "selected_model_id": "muse-spark-1.2-xhigh",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "qwen3.7-max",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": 0.012087533779340973,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.3",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh",
+            "qwen3.7-max",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.9468668617803175,
+          "selected_model_id": "deepseek-v4-pro-0813",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "task_completion_speed": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 36630585.75418182,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "quality_floor": 0.5792130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "gpt-5.6-sol-max",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "gpt-5.6-sol-max",
+            "gpt-5.5-xhigh",
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max"
+          ],
+          "quality_floor": 0.8197278750259673,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 9369974.166258065,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-max-effort",
+            "gemini-3.1-pro-preview-high",
+            "muse-spark-1.2-xhigh",
+            "muse-spark-1.3-xhigh",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.7425155351503601,
+          "selected_model_id": "claude-fable-5-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.231",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "deepseek-v4-pro-0813",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "deepseek-v4-pro-0813",
+            "deepseek-v4-pro",
+            "grok-4.3",
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "gpt-5.2-2025-12-11-high",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.6-flash-high",
+            "claude-opus-5-max-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "qwen3.7-max",
+            "muse-spark-1.3-xhigh",
+            "gemini-3.5-flash-high",
+            "kimi-k3",
+            "gemini-3.8-flash-high",
+            "gpt-5.5-xhigh",
+            "gpt-5.4-xhigh",
+            "qwen3.8-max",
+            "glm-5.3",
+            "gemini-3.1-pro-preview-high",
+            "claude-sonnet-5-xhigh-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-opus-4-5-20251101-thinking-64k-high-effort",
+            "claude-opus-4-6-thinking-auto-high-effort",
+            "claude-opus-4-7-xhigh-effort",
+            "claude-opus-4-8-max-effort",
+            "claude-opus-5-max-effort",
+            "claude-sonnet-5-xhigh-effort",
+            "deepseek-v4-pro",
+            "deepseek-v4-pro-0813",
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.5-flash-high",
+            "gemini-3.6-flash-high",
+            "gemini-3.7-flash-high",
+            "gemini-3.8-flash-high",
+            "glm-5.3",
+            "gpt-5.2-2025-12-11-high",
+            "gpt-5.4-xhigh",
+            "gpt-5.5-xhigh",
+            "gpt-5.6-sol-max",
+            "grok-4.3",
+            "grok-4.5",
+            "grok-4.6",
+            "kimi-k3",
+            "muse-spark-1.3-xhigh",
+            "qwen3.7-max",
+            "qwen3.8-max"
+          ],
+          "quality_floor": 0.9468668617803175,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      }
+    },
+    "max": {
+      "balanced": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 0.0,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "quality_floor": 0.5992130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "claude-fable-5-1-max-effort",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "quality_floor": 0.8397278750259674,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "fallback_no_speed_evidence",
+          "fallback_reason": "No quality-admitted model has Terminal-Bench completion-time evidence.",
+          "objective_eligible_model_ids": [
+            "muse-spark-1.3-xhigh"
+          ],
+          "objective_value": 0.236816088515293,
+          "quality_admitted_model_ids": [
+            "muse-spark-1.3-xhigh"
+          ],
+          "quality_floor": 0.7625155351503601,
+          "selected_model_id": "muse-spark-1.3-xhigh",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "gemini-3.7-flash-high",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.1-pro-preview-high"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-sol-max",
+            "grok-4.5",
+            "grok-4.6"
+          ],
+          "quality_floor": 0.9668668617803176,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "cost": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 4.002891945931923,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "quality_floor": 0.5992130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": 0.38505754057873887,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "quality_floor": 0.8397278750259674,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "muse-spark-1.3-xhigh"
+          ],
+          "objective_value": 0.236816088515293,
+          "quality_admitted_model_ids": [
+            "muse-spark-1.3-xhigh"
+          ],
+          "quality_floor": 0.7625155351503601,
+          "selected_model_id": "muse-spark-1.3-xhigh",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": null,
+          "evidence_status": "direct",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.1-pro-preview-high"
+          ],
+          "objective_value": 0.024636455346966592,
+          "quality_admitted_model_ids": [
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-sol-max",
+            "grok-4.5",
+            "grok-4.6"
+          ],
+          "quality_floor": 0.9668668617803176,
+          "selected_model_id": "gemini-3.7-flash-high",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      },
+      "task_completion_speed": {
+        "complex": {
+          "cold_start_model_id": null,
+          "evidence_status": "cohort_scoped",
+          "fallback_reason": null,
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "objective_value": 36630585.75418182,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort"
+          ],
+          "quality_floor": 0.5992130921684613,
+          "selected_model_id": "claude-fable-5-1-max-effort",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": "claude-code@2.1.257",
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "simple": {
+          "cold_start_model_id": "claude-fable-5-1-max-effort",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "claude-fable-5-1-max-effort",
+            "claude-fable-5-max-effort"
+          ],
+          "quality_floor": 0.8397278750259674,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        },
+        "standard": {
+          "cold_start_model_id": null,
+          "evidence_status": "fallback_no_speed_evidence",
+          "fallback_reason": "No quality-admitted model has Terminal-Bench completion-time evidence.",
+          "objective_eligible_model_ids": [
+            "muse-spark-1.3-xhigh"
+          ],
+          "objective_value": 0.236816088515293,
+          "quality_admitted_model_ids": [
+            "muse-spark-1.3-xhigh"
+          ],
+          "quality_floor": 0.7625155351503601,
+          "selected_model_id": "muse-spark-1.3-xhigh",
+          "selection_mode": "snapshot_ranked",
+          "speed_cohort_id": null,
+          "speed_metric": "offline_task_completion_latency"
+        },
+        "trivial": {
+          "cold_start_model_id": "gemini-3.7-flash-high",
+          "evidence_status": "fallback_runtime_response_speed_required",
+          "fallback_reason": "The offline snapshot intentionally has no provider-neutral response-latency value. LiteLLM must rerank these quality-admitted models using live TTFT and output throughput.",
+          "objective_eligible_model_ids": [
+            "gemini-3.7-flash-high",
+            "grok-4.6",
+            "grok-4.5",
+            "gpt-5.6-sol-max",
+            "gemini-3.1-pro-preview-high"
+          ],
+          "objective_value": null,
+          "quality_admitted_model_ids": [
+            "gemini-3.1-pro-preview-high",
+            "gemini-3.7-flash-high",
+            "gpt-5.6-sol-max",
+            "grok-4.5",
+            "grok-4.6"
+          ],
+          "quality_floor": 0.9668668617803176,
+          "selected_model_id": null,
+          "selection_mode": "runtime_response_latency",
+          "speed_cohort_id": null,
+          "speed_metric": "online_response_latency"
+        }
+      }
+    }
+  },
+  "schema_version": "2.1",
   "snapshot_id": "2026-09-03-v0",
   "task_complexities": {
     "livebench-2026-06-25": {
@@ -4573,6 +8165,74 @@
       "typescript": "complex",
       "typos": "simple",
       "zebra_puzzle": "simple"
+    },
+    "terminal-bench-4.0.0": {
+      "terminal-bench/atrx-vep-crispr": "trivial",
+      "terminal-bench/batched-eval-parity": "simple",
+      "terminal-bench/biped-contact-dynamics": "simple",
+      "terminal-bench/bun-sourcemap-leak": "complex",
+      "terminal-bench/cad-model": "simple",
+      "terminal-bench/cargo-flight-dispatch": "complex",
+      "terminal-bench/coq-block-bound": "trivial",
+      "terminal-bench/ctr-optimization": "complex",
+      "terminal-bench/cumulative-layout-shift": "trivial",
+      "terminal-bench/data-anonymization": "complex",
+      "terminal-bench/distributed-dedup": "simple",
+      "terminal-bench/embedding-drift-monitor": "simple",
+      "terminal-bench/fin-saccr-rwa": "trivial",
+      "terminal-bench/foodstuff-beta-activity": "complex",
+      "terminal-bench/formal-crypto": "standard",
+      "terminal-bench/fp8-rmsnorm-gemm": "trivial",
+      "terminal-bench/freecad-impeller": "complex",
+      "terminal-bench/freecad-platform-drawing": "standard",
+      "terminal-bench/freecad-spring-clip": "simple",
+      "terminal-bench/freight-dispatch-shift": "complex",
+      "terminal-bench/glycan-ms2-elucidation": "complex",
+      "terminal-bench/gsea-proteomics": "standard",
+      "terminal-bench/heat-pump-warranty": "standard",
+      "terminal-bench/hof-topology-interpenetration": "simple",
+      "terminal-bench/html-js-filter": "standard",
+      "terminal-bench/interleaved-vigenere": "simple",
+      "terminal-bench/intrastat-meldung": "trivial",
+      "terminal-bench/jax-speedrun-gpu": "complex",
+      "terminal-bench/ks-solver-cpp": "standard",
+      "terminal-bench/kv-live-surgery": "simple",
+      "terminal-bench/lake-temp-glm": "complex",
+      "terminal-bench/layout-config-recreation": "complex",
+      "terminal-bench/layout-config-recreation2": "trivial",
+      "terminal-bench/legacy-utility-triage": "standard",
+      "terminal-bench/live-database-cutover": "standard",
+      "terminal-bench/math-eval-grader": "standard",
+      "terminal-bench/medical-claims-processing": "complex",
+      "terminal-bench/mp-checkpoint-consolidation": "trivial",
+      "terminal-bench/music-harmony": "complex",
+      "terminal-bench/mvcc-lsm-compaction": "simple",
+      "terminal-bench/nextjs-performance": "standard",
+      "terminal-bench/ontology-kg-querying": "complex",
+      "terminal-bench/payments-pipeline-fix": "simple",
+      "terminal-bench/photonic-waveguide-routing": "simple",
+      "terminal-bench/pretrain-shard-corruption": "standard",
+      "terminal-bench/production-planning": "standard",
+      "terminal-bench/protein-autointerp-disulfide": "complex",
+      "terminal-bench/react-lead-form": "trivial",
+      "terminal-bench/retro-console-soc": "trivial",
+      "terminal-bench/risk-scorer-replay": "trivial",
+      "terminal-bench/roy-polymorph-cn": "complex",
+      "terminal-bench/rs-archive-clone": "simple",
+      "terminal-bench/satb-audio-transcription": "standard",
+      "terminal-bench/session-window-debug": "complex",
+      "terminal-bench/sglang-qwen-burst": "standard",
+      "terminal-bench/shadow-relay": "trivial",
+      "terminal-bench/sound-change-cascade": "simple",
+      "terminal-bench/takens-embedding-lean": "standard",
+      "terminal-bench/telecom-entity-resolution": "trivial",
+      "terminal-bench/uefi-bootkit": "trivial",
+      "terminal-bench/vba-userform-port": "simple",
+      "terminal-bench/vf2-speedup-networkx": "simple",
+      "terminal-bench/vllm-deepseek-streaming": "standard",
+      "terminal-bench/vpp-loss-divergence": "trivial",
+      "terminal-bench/wal-recovery-ordering": "simple",
+      "terminal-bench/wdm-design": "trivial"
     }
   }
 }
diff --git a/litellm/router_strategy/complexity_router/auto_setup.py b/litellm/router_strategy/complexity_router/auto_setup.py
index 9e41a4d25d8..2c61cf1aa0f 100644
--- a/litellm/router_strategy/complexity_router/auto_setup.py
+++ b/litellm/router_strategy/complexity_router/auto_setup.py
@@ -13,10 +13,16 @@ from typing import Final, Literal
 
 from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, model_validator
 
-from .config import ComplexityRouterConfig
+from .config import (
+    AutoSetupCandidate,
+    AutoSetupConfig,
+    AutoSetupObjective,
+    AutoSetupQualityLevel,
+    AutoSetupTierPolicy,
+    ComplexityRouterConfig,
+)
 
 SnapshotComplexity = Literal["trivial", "simple", "standard", "complex"]
-AutoSetupQualityLevel = Literal["economy", "balanced", "high", "max"]
 
 _SNAPSHOT_FILENAME: Final = "auto_router_snapshot_v0.json"
 _TIER_COMPLEXITIES: Final[Mapping[str, SnapshotComplexity]] = MappingProxyType(
@@ -27,6 +33,7 @@ _TIER_COMPLEXITIES: Final[Mapping[str, SnapshotComplexity]] = MappingProxyType(
         "REASONING": "complex",
     }
 )
+_EASY_COMPLEXITIES: Final = frozenset(("trivial", "simple"))
 _VERSION_SEPARATOR: Final = re.compile(r"(?<=\d)\.(?=\d)")
 _SNAPSHOT_PAYLOAD_ADAPTER: Final = TypeAdapter(dict[str, object])
 
@@ -68,12 +75,36 @@ class SnapshotQualityProfile(BaseModel):
     subtask_count: int = Field(gt=0)
 
 
+class SnapshotSpeedProfile(BaseModel):
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    attempt_count: int = Field(gt=0)
+    benchmark_model_id: str
+    cohort_id: str
+    complexity: SnapshotComplexity
+    cost_per_completed_task_usd: float | None = Field(default=None, ge=0)
+    excluded_infrastructure_trials: int = Field(ge=0)
+    mean_attempt_cost_usd: float | None = Field(default=None, ge=0)
+    mean_cache_read_input_tokens: float | None = Field(default=None, ge=0)
+    mean_output_tokens: float | None = Field(default=None, ge=0)
+    mean_uncached_input_tokens: float | None = Field(default=None, ge=0)
+    observed_duration_p50_ms: float = Field(ge=0)
+    observed_duration_p95_ms: float = Field(ge=0)
+    quality_lower_bound: float = Field(ge=0, le=1)
+    retry_adjusted_completion_ms: float | None = Field(default=None, ge=0)
+    source_id: str
+    success_probability: float = Field(ge=0, le=1)
+    task_count: int = Field(gt=0)
+    terminal_bench_model_id: str
+
+
 class SnapshotModel(BaseModel):
     model_config = ConfigDict(extra="forbid", frozen=True)
 
     benchmark_model_id: str
     identity: SnapshotIdentity
     quality_by_complexity: Mapping[SnapshotComplexity, SnapshotQualityProfile]
+    task_completion_speed_by_complexity: Mapping[SnapshotComplexity, tuple[SnapshotSpeedProfile, ...]]
 
 
 class AutoRouterSnapshot(BaseModel):
@@ -88,7 +119,8 @@ class AutoRouterSnapshot(BaseModel):
     models: Mapping[str, SnapshotModel]
     provenance: Mapping[str, object]
     quality_tiers: Mapping[AutoSetupQualityLevel, SnapshotQualityTier]
-    schema_version: Literal["3.0"]
+    reference_routes: Mapping[str, object]
+    schema_version: Literal["2.1"]
     snapshot_id: str
     task_complexities: Mapping[str, Mapping[str, SnapshotComplexity]]
 
@@ -105,12 +137,24 @@ class AutoRouterSnapshot(BaseModel):
                 raise ValueError(f"Auto Router snapshot model identity mismatch for {model_id}")
             if frozenset(model.quality_by_complexity) != expected_complexities:
                 raise ValueError(f"Auto Router snapshot model {model_id} lacks complete quality evidence")
+            if frozenset(model.task_completion_speed_by_complexity) != expected_complexities:
+                raise ValueError(f"Auto Router snapshot model {model_id} lacks complete speed buckets")
             if model.identity.is_routable != bool(model.identity.litellm_model_keys):
                 raise ValueError(f"Auto Router snapshot routability mismatch for {model_id}")
             for complexity in expected_complexities:
                 quality = model.quality_by_complexity[complexity]
                 if quality.benchmark_model_id != model_id or quality.complexity != complexity:
                     raise ValueError(f"Auto Router snapshot quality identity mismatch for {model_id}/{complexity}")
+                for speed in model.task_completion_speed_by_complexity[complexity]:
+                    token_values = (
+                        speed.mean_uncached_input_tokens,
+                        speed.mean_cache_read_input_tokens,
+                        speed.mean_output_tokens,
+                    )
+                    if any(value is None for value in token_values) != all(value is None for value in token_values):
+                        raise ValueError(f"Auto Router snapshot has partial token evidence for {model_id}/{complexity}")
+                    if speed.benchmark_model_id != model_id or speed.complexity != complexity:
+                        raise ValueError(f"Auto Router snapshot speed identity mismatch for {model_id}/{complexity}")
             for alias in model.identity.litellm_model_keys:
                 normalized = _normalized_model_key(alias)
                 owner = seen_aliases.setdefault(normalized, model.identity.family_id)
@@ -149,7 +193,7 @@ class AutoSetupDeployment(BaseModel):
     pricing: AutoSetupDeploymentPricing | None = None
 
 
-AutoSetupExclusionReason = Literal["no_benchmark_match", "mixed_model_group"]
+AutoSetupExclusionReason = Literal["no_benchmark_match", "mixed_model_group", "pricing_unavailable"]
 
 
 @dataclass(frozen=True, slots=True)
@@ -162,8 +206,9 @@ class AutoSetupInventoryExclusion:
 class _AvailableProfile:
     model_name: str
     benchmark_model_id: str
-    cost_per_completed_task_usd: float | None
+    cost_per_completed_task_usd: float
     quality: SnapshotQualityProfile
+    speed: tuple[SnapshotSpeedProfile, ...]
     required_parameters: Mapping[str, object]
 
 
@@ -253,7 +298,18 @@ def analyze_auto_setup_inventory(
     """Return identity-safe groups and explicit reasons for every excluded group."""
 
     groups, exclusions = _resolve_groups(snapshot, available_model_deployments)
-    return tuple(group.model_name for group in groups), exclusions
+    eligible: list[str] = []
+    mutable_exclusions = list(exclusions)
+    for group in groups:
+        if any(
+            _quality_completion_cost(model.quality_by_complexity[complexity], group.deployments) is not None
+            for model in group.matches
+            for complexity in _TIER_COMPLEXITIES.values()
+        ):
+            eligible.append(group.model_name)
+        else:
+            mutable_exclusions.append(AutoSetupInventoryExclusion(group.model_name, "pricing_unavailable"))
+    return tuple(eligible), tuple(mutable_exclusions)
 
 
 def _rate_for_input_size(
@@ -289,6 +345,19 @@ def _rate_for_input_size(
     return base
 
 
+def _base_rate(
+    pricing: AutoSetupDeploymentPricing,
+    field: Literal["input", "output", "cache"],
+) -> float:
+    if field == "input":
+        return pricing.input_cost_per_token
+    if field == "output":
+        return pricing.output_cost_per_token
+    if pricing.cache_read_input_token_cost is not None:
+        return pricing.cache_read_input_token_cost
+    return pricing.input_cost_per_token
+
+
 def _quality_completion_cost(
     profile: SnapshotQualityProfile,
     deployments: Sequence[AutoSetupDeployment],
@@ -312,6 +381,31 @@ def _quality_completion_cost(
     return max(costs, default=None)
 
 
+def _speed_completion_cost(
+    speed: SnapshotSpeedProfile,
+    deployments: Sequence[AutoSetupDeployment],
+) -> float | None:
+    if (
+        speed.success_probability <= 0
+        or speed.mean_uncached_input_tokens is None
+        or speed.mean_cache_read_input_tokens is None
+        or speed.mean_output_tokens is None
+    ):
+        return None
+    costs: list[float] = []
+    for deployment in deployments:
+        pricing = deployment.pricing
+        if pricing is None:
+            return None
+        attempt_cost = (
+            speed.mean_uncached_input_tokens * _base_rate(pricing, "input")
+            + speed.mean_cache_read_input_tokens * _base_rate(pricing, "cache")
+            + speed.mean_output_tokens * _base_rate(pricing, "output")
+        )
+        costs.append(attempt_cost / speed.success_probability)
+    return max(costs, default=None)
+
+
 def _available_profiles(
     snapshot: AutoRouterSnapshot,
     available_model_deployments: Mapping[str, Sequence[AutoSetupDeployment]],
@@ -333,12 +427,19 @@ def _available_profiles(
             ),
         )
         cost = _quality_completion_cost(best.quality_by_complexity[complexity], group.deployments)
+        if cost is None:
+            continue
+        priced_speed = tuple(
+            speed.model_copy(update={"cost_per_completed_task_usd": _speed_completion_cost(speed, group.deployments)})
+            for speed in best.task_completion_speed_by_complexity[complexity]
+        )
         profiles.append(
             _AvailableProfile(
                 model_name=group.model_name,
                 benchmark_model_id=best.benchmark_model_id,
                 cost_per_completed_task_usd=cost,
                 quality=best.quality_by_complexity[complexity],
+                speed=priced_speed,
                 required_parameters=best.identity.required_parameters,
             )
         )
@@ -350,8 +451,7 @@ def _cost_rank(profiles: Sequence[_AvailableProfile]) -> tuple[_AvailableProfile
         sorted(
             profiles,
             key=lambda profile: (
-                profile.cost_per_completed_task_usd is None,
-                profile.cost_per_completed_task_usd if profile.cost_per_completed_task_usd is not None else math.inf,
+                profile.cost_per_completed_task_usd,
                 -profile.quality.quality_lower_bound,
                 profile.model_name,
             ),
@@ -359,15 +459,146 @@ def _cost_rank(profiles: Sequence[_AvailableProfile]) -> tuple[_AvailableProfile
     )
 
 
+def _normalize_log(value: float, values: Sequence[float]) -> float:
+    logged: Final = tuple(math.log(max(item, 1e-12)) for item in values)
+    low: Final = min(logged)
+    high: Final = max(logged)
+    if low == high:
+        return 0.0
+    return (math.log(max(value, 1e-12)) - low) / (high - low)
+
+
+def _hard_task_rank(
+    profiles: Sequence[_AvailableProfile], objective: AutoSetupObjective
+) -> tuple[tuple[_AvailableProfile, ...], str]:
+    cost_ranked: Final = _cost_rank(profiles)
+    speed_profiles: Final = tuple(
+        (profile, speed)
+        for profile in profiles
+        for speed in profile.speed
+        if speed.retry_adjusted_completion_ms is not None
+    )
+    if not speed_profiles:
+        return cost_ranked, "fallback_no_speed_evidence"
+    cohorts: dict[  # mutable-ok: local cohort index never escapes
+        str, dict[str, tuple[_AvailableProfile, SnapshotSpeedProfile]]
+    ] = {}  # mutable-ok: local cohort index is populated before selection
+    for profile, speed in speed_profiles:
+        by_model = cohorts.setdefault(
+            speed.cohort_id,
+            {},  # mutable-ok: local cohort index needs an empty per-cohort bucket
+        )
+        current = by_model.get(profile.model_name)
+        if current is None or speed.attempt_count > current[1].attempt_count:
+            by_model[profile.model_name] = (profile, speed)
+    measured_profiles: Final = tuple(
+        {  # mutable-ok: transient dict provides deterministic model-name deduplication
+            profile.model_name: profile for profile, _ in speed_profiles
+        }.values()
+    )
+    anchor: Final = max(
+        measured_profiles,
+        key=lambda profile: (profile.quality.quality_lower_bound, -profile.cost_per_completed_task_usd),
+    )
+    anchor_cohorts: Final = tuple(cohort_id for cohort_id, by_model in cohorts.items() if anchor.model_name in by_model)
+    selected_cohort = max(
+        anchor_cohorts,
+        key=lambda cohort_id: (
+            len(cohorts[cohort_id]),
+            sum(speed.attempt_count for _, speed in cohorts[cohort_id].values()),
+            cohort_id,
+        ),
+    )
+    comparable = tuple(cohorts[selected_cohort].values())
+    if objective == "task_completion_speed":
+        return tuple(
+            profile
+            for profile, _ in sorted(
+                comparable,
+                key=lambda item: (
+                    item[1].retry_adjusted_completion_ms or math.inf,
+                    -item[0].quality.quality_lower_bound,
+                    item[0].model_name,
+                ),
+            )
+        ), "cohort_scoped"
+    balanced_comparable: list[  # mutable-ok: local typed accumulator is consumed immediately
+        tuple[_AvailableProfile, float, float]
+    ] = []  # mutable-ok: local typed accumulator is frozen by the return expression
+    for profile, speed in comparable:
+        cost = speed.cost_per_completed_task_usd
+        completion_ms = speed.retry_adjusted_completion_ms
+        if cost is not None and completion_ms is not None:
+            balanced_comparable.append((profile, cost, completion_ms))
+    if not balanced_comparable:
+        return cost_ranked, "fallback_no_comparable_cost_evidence"
+    costs: Final = tuple(cost for _, cost, _ in balanced_comparable)
+    times: Final = tuple(completion_ms for _, _, completion_ms in balanced_comparable)
+    return tuple(
+        profile
+        for profile, _, _ in sorted(
+            balanced_comparable,
+            key=lambda item: (
+                math.hypot(
+                    _normalize_log(item[1], costs),
+                    _normalize_log(item[2], times),
+                ),
+                -item[0].quality.quality_lower_bound,
+                item[0].model_name,
+            ),
+        )
+    ), "cohort_scoped"
+
+
+def _policy_candidate(profile: _AvailableProfile) -> AutoSetupCandidate:
+    return AutoSetupCandidate(
+        model_name=profile.model_name,
+        benchmark_model_id=profile.benchmark_model_id,
+        quality_lower_bound=profile.quality.quality_lower_bound,
+        cost_per_completed_task_usd=profile.cost_per_completed_task_usd,
+    )
+
+
+def _tier_policy(
+    profiles: Sequence[_AvailableProfile],
+    complexity: SnapshotComplexity,
+    objective: AutoSetupObjective,
+) -> AutoSetupTierPolicy:
+    cost_ranked: Final = _cost_rank(profiles)
+    if objective == "cost":
+        return AutoSetupTierPolicy(
+            selection_mode="snapshot_ranked",
+            candidates=tuple(_policy_candidate(profile) for profile in cost_ranked),
+            evidence_status="direct",
+        )
+    if complexity in _EASY_COMPLEXITIES:
+        return AutoSetupTierPolicy(
+            selection_mode="runtime_response_latency",
+            candidates=tuple(_policy_candidate(profile) for profile in cost_ranked),
+            cold_start_model=cost_ranked[0].model_name,
+            evidence_status="runtime_response_speed_required",
+        )
+    ranked, evidence_status = _hard_task_rank(profiles, objective)
+    return AutoSetupTierPolicy(
+        selection_mode="snapshot_ranked",
+        candidates=tuple(_policy_candidate(profile) for profile in ranked),
+        evidence_status=evidence_status,
+    )
+
+
 def build_auto_setup_config(
     *,
     snapshot: AutoRouterSnapshot,
     available_model_deployments: Mapping[str, Sequence[AutoSetupDeployment]],
     quality_level: AutoSetupQualityLevel,
+    optimize_for: AutoSetupObjective,
 ) -> ComplexityRouterConfig:
     """Recompute quality gates and rankings over only the caller's available groups."""
 
     maximum_regret: Final = snapshot.quality_tiers[quality_level].maximum_quality_regret
+    policies: Final[  # mutable-ok: Pydantic config assembly is local to this builder
+        dict[str, AutoSetupTierPolicy]
+    ] = {}  # mutable-ok: Pydantic config assembly is local to this builder
     tiers: Final[dict[str, list[str]]] = {}  # mutable-ok: Pydantic config assembly is local to this builder
     tier_model_configs: Final[  # mutable-ok: Pydantic config assembly is local to this builder
         dict[str, list[dict[str, object]]]
@@ -379,20 +610,21 @@ def build_auto_setup_config(
         best_quality = max(profile.quality.quality_lower_bound for profile in profiles)
         floor = max(0.0, best_quality - maximum_regret)
         admitted = tuple(profile for profile in profiles if profile.quality.quality_lower_bound >= floor)
-        selected = _cost_rank(admitted)[0]
-        tiers[tier_name] = [selected.model_name]
-        parameterized: list[dict[str, object]] = (
-            [  # mutable-ok: Pydantic validates and owns these JSON-shaped rows
-                {  # mutable-ok: each row is a Pydantic input payload
-                    "model_name": selected.model_name,
-                    "litellm_params": dict(  # mutable-ok: Pydantic needs a concrete JSON-shaped params object
-                        selected.required_parameters
-                    ),
-                }
-            ]
-            if selected.required_parameters
-            else []
-        )
+        policy = _tier_policy(admitted, complexity, optimize_for)
+        policies[tier_name] = policy
+        tiers[tier_name] = [  # mutable-ok: ComplexityRouterConfig owns this JSON-shaped tier list
+            candidate.model_name for candidate in policy.candidates
+        ]
+        parameterized: list[dict[str, object]] = [  # mutable-ok: Pydantic validates and owns these JSON-shaped rows
+            {  # mutable-ok: each row is a Pydantic input payload
+                "model_name": profile.model_name,
+                "litellm_params": dict(  # mutable-ok: Pydantic needs a concrete JSON-shaped params object
+                    profile.required_parameters
+                ),
+            }
+            for profile in admitted
+            if profile.model_name in frozenset(tiers[tier_name]) and profile.required_parameters
+        ]
         if parameterized:
             tier_model_configs[tier_name] = parameterized
 
@@ -401,5 +633,12 @@ def build_auto_setup_config(
             "tiers": tiers,
             "tier_model_configs": tier_model_configs,
             "classifier_type": "heuristic_v2",
+            "auto_setup": AutoSetupConfig(
+                snapshot_id=snapshot.snapshot_id,
+                snapshot_sha256=snapshot.artifact_sha256,
+                quality_level=quality_level,
+                optimize_for=optimize_for,
+                tier_policies=policies,
+            ).model_dump(mode="json"),
         }
     )
diff --git a/litellm/router_strategy/complexity_router/complexity_router.py b/litellm/router_strategy/complexity_router/complexity_router.py
index 17e3d1256d0..8f06b0c97f4 100644
--- a/litellm/router_strategy/complexity_router/complexity_router.py
+++ b/litellm/router_strategy/complexity_router/complexity_router.py
@@ -1322,6 +1322,13 @@ class ComplexityRouter(CustomLogger):
         if tier is not None:
             tier_name: Final = _tier_name(tier)
             decision["tier"] = tier_name
+            if self.config.auto_setup is not None:
+                auto_policy: Final = self.config.auto_setup.tier_policies.get(tier_name)
+                if auto_policy is not None:
+                    decision["auto_setup_snapshot_id"] = self.config.auto_setup.snapshot_id
+                    decision["auto_setup_quality_level"] = self.config.auto_setup.quality_level
+                    decision["auto_setup_optimize_for"] = self.config.auto_setup.optimize_for
+                    decision["auto_setup_selection_mode"] = auto_policy.selection_mode
             if not self.config.has_custom_tiers:
                 label = self.config.tier_label(ComplexityTier(tier_name))
                 if label != tier_name:
@@ -1815,6 +1822,37 @@ class ComplexityRouter(CustomLogger):
             raise ValueError(f"Empty model pool for tier {tier_key}")
         return random.choice(model)
 
+    async def _pick_from_auto_setup_policy(self, tier_key: str, pool: Sequence[str]) -> str | None:
+        setup: Final = self.config.auto_setup
+        if setup is None:
+            return None
+        policy: Final = setup.tier_policies.get(tier_key)
+        if policy is None:
+            return None
+        available: Final = tuple(dict.fromkeys(pool))
+        if not available:
+            raise ValueError(f"No models configured for tier {tier_key}")
+        ranked: Final = tuple(
+            candidate.model_name for candidate in policy.candidates if candidate.model_name in frozenset(available)
+        )
+        if policy.selection_mode == "snapshot_ranked":
+            return ranked[0] if ranked else available[0]
+
+        from .response_latency import select_runtime_response_model
+
+        cold_start: Final = policy.cold_start_model
+        if cold_start is None:
+            raise ValueError(f"Auto setup runtime response-latency policy for {tier_key} has no cold-start model")
+        return await select_runtime_response_model(
+            router_cache=self.litellm_router_instance.cache,
+            router_model_name=self.model_name,
+            tier=tier_key,
+            candidates=policy.candidates,
+            available_models=available,
+            cold_start_model=cold_start,
+            objective=setup.optimize_for,
+        )
+
     def _tier_pools(self) -> dict[str, list[str]]:
         return {tier: (models if isinstance(models, list) else [models]) for tier, models in self.config.tiers.items()}
 
@@ -1826,17 +1864,25 @@ class ComplexityRouter(CustomLogger):
         request_kwargs: dict,
         allowed_models: tuple[str, ...] | None = None,
     ) -> str:
+        tier_key = _tier_name(tier)
         if not self.config.plugins:
-            if allowed_models is not None:
-                return self._pick_from_tier_value(allowed_models, _tier_name(tier))
-            return self.get_model_for_tier(tier)
+            configured = self.config.tiers.get(tier_key)
+            if configured is None:
+                return self.get_model_for_tier(tier)
+            full_pool = (configured,) if isinstance(configured, str) else tuple(configured)
+            pool = (
+                tuple(model for model in full_pool if model in allowed_models)
+                if allowed_models is not None
+                else full_pool
+            )
+            direct_auto_pick: Final = await self._pick_from_auto_setup_policy(tier_key, pool)
+            return direct_auto_pick if direct_auto_pick is not None else self._pick_from_tier_value(pool, tier_key)
 
         from litellm.types.router import RoutingContext
 
-        tier_key: Final = _tier_name(tier)
         metadata_key: Final = get_metadata_variable_name_from_kwargs(request_kwargs)
-        full_pool: Final = tuple(self._tier_pools().get(tier_key, ()))
-        pool: Final = (
+        full_pool = tuple(self._tier_pools().get(tier_key, ()))
+        pool = (
             tuple(model for model in full_pool if model in allowed_models) if allowed_models is not None else full_pool
         )
         if not pool:
@@ -1860,7 +1906,12 @@ class ComplexityRouter(CustomLogger):
             # silently bypassed. Raise instead, matching the Router-level plugin
             # pipeline's own fail-closed behavior for the same situation.
             raise ValueError(f"No candidate models left for tier {tier_key} after routing-plugin filtering")
-        return self._pick_from_tier_value(context.candidate_models, tier_key)
+        filtered_auto_pick: Final = await self._pick_from_auto_setup_policy(tier_key, context.candidate_models)
+        return (
+            filtered_auto_pick
+            if filtered_auto_pick is not None
+            else self._pick_from_tier_value(context.candidate_models, tier_key)
+        )
 
     def _ensure_adaptive_router(self) -> Any | None:
         if not self.config.adaptive:
diff --git a/litellm/router_strategy/complexity_router/config.py b/litellm/router_strategy/complexity_router/config.py
index 70c1b281e31..e1f78e3d969 100644
--- a/litellm/router_strategy/complexity_router/config.py
+++ b/litellm/router_strategy/complexity_router/config.py
@@ -27,6 +27,81 @@ class ComplexityTier(str, Enum):
     REASONING = "REASONING"
 
 
+AutoSetupQualityLevel = Literal["economy", "balanced", "high", "max"]
+AutoSetupObjective = Literal["cost", "task_completion_speed", "balanced"]
+AutoSetupSelectionMode = Literal["snapshot_ranked", "runtime_response_latency"]
+
+
+class AutoSetupCandidate(BaseModel):
+    """One benchmark-backed model group in an Auto-generated tier policy."""
+
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    model_name: str = Field(min_length=1)
+    benchmark_model_id: str = Field(min_length=1)
+    quality_lower_bound: float = Field(ge=0, le=1)
+    cost_per_completed_task_usd: float = Field(ge=0)
+
+
+class AutoSetupTierPolicy(BaseModel):
+    """How an Auto-generated tier chooses from its quality-admitted model groups."""
+
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    selection_mode: AutoSetupSelectionMode
+    candidates: tuple[AutoSetupCandidate, ...] = Field(min_length=1)
+    cold_start_model: str | None = None
+    evidence_status: str = Field(min_length=1)
+
+    @model_validator(mode="after")
+    def _validate_candidates(self) -> "AutoSetupTierPolicy":
+        names: Final = tuple(candidate.model_name for candidate in self.candidates)
+        if len(names) != len(frozenset(names)):
+            raise ValueError("Auto setup tier candidates must use distinct model_name values")
+        if self.selection_mode == "runtime_response_latency":
+            if self.cold_start_model not in names:
+                raise ValueError("A runtime response-latency policy needs a cold_start_model in candidates")
+        elif self.cold_start_model is not None:
+            raise ValueError("cold_start_model is only valid for runtime_response_latency")
+        return self
+
+
+class AutoSetupConfig(BaseModel):
+    """Provenance and selection policy stamped only by zero-setup Auto creation."""
+
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    schema_version: Literal["1"] = "1"
+    snapshot_id: str = Field(min_length=1)
+    snapshot_sha256: str = Field(pattern=r"^[0-9a-f]{64}$")
+    quality_level: AutoSetupQualityLevel
+    optimize_for: AutoSetupObjective
+    tier_policies: Mapping[str, AutoSetupTierPolicy]
+
+    @field_validator("tier_policies", mode="before")
+    @classmethod
+    def _freeze_tier_policies(cls, value: Mapping[str, object]) -> Mapping[str, object]:
+        return MappingProxyType(dict(value))
+
+    @field_serializer("tier_policies")
+    def _serialize_tier_policies(self, value: Mapping[str, AutoSetupTierPolicy]) -> Mapping[str, AutoSetupTierPolicy]:
+        return dict(value)  # mutable-ok: Pydantic JSON serialization requires a concrete mapping
+
+    @model_validator(mode="after")
+    def _validate_tier_policies(self) -> "AutoSetupConfig":
+        expected: Final = frozenset(tier.value for tier in TIER_SEVERITY_ORDER)
+        actual: Final = frozenset(self.tier_policies)
+        if actual != expected:
+            raise ValueError(
+                f"Auto setup tier_policies must define exactly {', '.join(tier.value for tier in TIER_SEVERITY_ORDER)}"
+            )
+        if self.optimize_for == "cost" and any(
+            policy.selection_mode != "snapshot_ranked" for policy in self.tier_policies.values()
+        ):
+            raise ValueError("Cost-optimized Auto setup cannot enable runtime response-latency selection")
+        return self
+
+
 class ClassificationRubric(str, Enum):
     """Which calibration examples, and for BUSINESS which tier criteria, the built-in classifier rubric carries."""
 
@@ -517,6 +592,14 @@ class ComplexityRouterConfig(BaseModel):
         default_factory=dict,
     )
 
+    auto_setup: AutoSetupConfig | None = Field(
+        default=None,
+        description=(
+            "Benchmark snapshot and per-tier selection policy generated by the zero-setup Auto flow. "
+            "Absent on manually configured routers, whose pool selection behavior is unchanged"
+        ),
+    )
+
     tier_definitions: tuple[TierDefinition, ...] | None = Field(
         default=None,
         description=(
@@ -1422,6 +1505,24 @@ class ComplexityRouterConfig(BaseModel):
             )
         return self
 
+    @model_validator(mode="after")
+    def _validate_auto_setup(self) -> "ComplexityRouterConfig":
+        setup: Final = self.auto_setup
+        if setup is None:
+            return self
+        if self.adaptive:
+            raise ValueError("auto_setup and adaptive=True cannot both select from the same tier pool")
+        if self.tier_definitions is not None:
+            raise ValueError("auto_setup requires the built-in four-tier complexity ladder")
+        for tier_name, policy in setup.tier_policies.items():
+            configured = self.tiers.get(tier_name)
+            configured_names = frozenset((configured,)) if isinstance(configured, str) else frozenset(configured or ())
+            policy_names = frozenset(candidate.model_name for candidate in policy.candidates)
+            if not policy_names.issubset(configured_names):
+                missing = ", ".join(sorted(policy_names - configured_names))
+                raise ValueError(f"auto_setup tier {tier_name} references models not present in tiers: {missing}")
+        return self
+
     @model_validator(mode="after")
     def _validate_tier_param_placement(self) -> "ComplexityRouterConfig":
         """Reject a router setting written into a tier entry's request params.
diff --git a/litellm/router_strategy/complexity_router/response_latency.py b/litellm/router_strategy/complexity_router/response_latency.py
new file mode 100644
index 00000000000..c77f9b46c23
--- /dev/null
+++ b/litellm/router_strategy/complexity_router/response_latency.py
@@ -0,0 +1,248 @@
+from __future__ import annotations
+
+import hashlib
+import math
+import statistics
+from collections.abc import Mapping, Sequence
+from datetime import datetime, timedelta
+from types import MappingProxyType
+from typing import Final, Protocol
+
+from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError
+
+from litellm._logging import verbose_router_logger
+
+from .config import AutoSetupCandidate, AutoSetupObjective
+
+_CACHE_PREFIX: Final = "auto_setup_response_latency:v1"
+_CACHE_TTL_SECONDS: Final = 7 * 24 * 60 * 60
+_MAX_SAMPLES_PER_MODEL: Final = 20
+_MAX_EXPECTED_OUTPUT_SAMPLES: Final = 100
+_MIN_SAMPLES_PER_MODEL: Final = 2
+_OBJECT_ADAPTER: Final = TypeAdapter(object)
+_OBJECT_MAPPING_ADAPTER: Final = TypeAdapter(dict[str, object])
+
+
+class ResponseLatencyCache(Protocol):
+    async def async_get_cache(
+        self,
+        key: str,
+        parent_otel_span: object | None = None,
+        local_only: bool = False,
+        **kwargs: object,  # kwargs-ok: matches the shared cache interface's optional transport parameters
+    ) -> object: ...
+
+    async def async_set_cache(
+        self,
+        key: str,
+        value: object,
+        local_only: bool = False,
+        **kwargs: object,  # kwargs-ok: matches the shared cache interface's optional transport parameters
+    ) -> None: ...
+
+
+class ResponseLatencySample(BaseModel):
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    ttft_ms: float = Field(ge=0)
+    output_tokens_per_ms: float = Field(gt=0)
+
+
+class ResponseLatencyState(BaseModel):
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    samples_by_model: Mapping[str, tuple[ResponseLatencySample, ...]] = Field(
+        default_factory=lambda: MappingProxyType({})
+    )
+    visible_output_tokens: tuple[int, ...] = ()
+
+
+def _cache_key(router_model_name: str, tier: str) -> str:
+    namespace: Final = hashlib.sha256(router_model_name.encode("utf-8")).hexdigest()[:24]
+    return f"{_CACHE_PREFIX}:{namespace}:{tier}"
+
+
+def _state(value: object) -> ResponseLatencyState:
+    try:
+        return ResponseLatencyState.model_validate(value or MappingProxyType({}))
+    except ValidationError:
+        return ResponseLatencyState()
+
+
+def _candidate_order(
+    candidates: Sequence[AutoSetupCandidate],
+    available_models: Sequence[str],
+    cold_start_model: str,
+) -> tuple[str, ...]:
+    available: Final = frozenset(available_models)
+    ranked: Final = tuple(candidate.model_name for candidate in candidates if candidate.model_name in available)
+    extras: Final = tuple(model for model in available_models if model not in frozenset(ranked))
+    ordered: Final = tuple(dict.fromkeys((*ranked, *extras)))
+    if cold_start_model not in ordered:
+        return ordered
+    return (cold_start_model, *(model for model in ordered if model != cold_start_model))
+
+
+def _median_prediction_ms(samples: Sequence[ResponseLatencySample], expected_output_tokens: float) -> float:
+    return statistics.median(
+        sample.ttft_ms + expected_output_tokens / sample.output_tokens_per_ms for sample in samples
+    )
+
+
+def _normalized_log(value: float, values: Sequence[float]) -> float:
+    logged: Final = tuple(math.log(max(item, 1e-12)) for item in values)
+    low: Final = min(logged)
+    high: Final = max(logged)
+    if low == high:
+        return 0.0
+    return (math.log(max(value, 1e-12)) - low) / (high - low)
+
+
+def _best_observed_model(
+    state: ResponseLatencyState,
+    candidates: Sequence[AutoSetupCandidate],
+    ordered_models: Sequence[str],
+    objective: AutoSetupObjective,
+) -> str:
+    expected_tokens: Final = statistics.median(state.visible_output_tokens) if state.visible_output_tokens else 1.0
+    latencies: Final[Mapping[str, float]] = MappingProxyType(
+        {model: _median_prediction_ms(state.samples_by_model[model], expected_tokens) for model in ordered_models}
+    )
+    if objective == "task_completion_speed":
+        return min(ordered_models, key=lambda model: (latencies[model], ordered_models.index(model)))
+
+    configured_costs: Final[Mapping[str, float]] = MappingProxyType(
+        {candidate.model_name: candidate.cost_per_completed_task_usd for candidate in candidates}
+    )
+    known_costs: Final = tuple(configured_costs.values())
+    conservative_unknown_cost: Final = max(known_costs) if known_costs else 1.0
+    costs: Final[Mapping[str, float]] = MappingProxyType(
+        {model: configured_costs.get(model, conservative_unknown_cost) for model in ordered_models}
+    )
+    latency_values: Final = tuple(latencies.values())
+    cost_values: Final = tuple(costs.values())
+    return min(
+        ordered_models,
+        key=lambda model: (
+            math.hypot(
+                _normalized_log(latencies[model], latency_values),
+                _normalized_log(costs[model], cost_values),
+            ),
+            ordered_models.index(model),
+        ),
+    )
+
+
+async def select_runtime_response_model(
+    *,
+    router_cache: ResponseLatencyCache,
+    router_model_name: str,
+    tier: str,
+    candidates: Sequence[AutoSetupCandidate],
+    available_models: Sequence[str],
+    cold_start_model: str,
+    objective: AutoSetupObjective,
+) -> str:
+    """Explore quality-admitted groups, then minimize equal-output response time."""
+
+    ordered: Final = _candidate_order(candidates, available_models, cold_start_model)
+    if not ordered:
+        raise ValueError(f"No Auto setup candidates remain for tier {tier}")
+    try:
+        cached = await router_cache.async_get_cache(key=_cache_key(router_model_name, tier))
+        state: Final = _state(cached)
+    except Exception as exc:  # noqa: BLE001 -- optional observations must never make a request unroutable
+        verbose_router_logger.warning("Auto setup response-latency cache read failed: %s", exc)
+        return ordered[0]
+
+    counts: Final[Mapping[str, int]] = MappingProxyType(
+        {model: len(state.samples_by_model.get(model, ())) for model in ordered}
+    )
+    least_observed: Final = min(counts.values())
+    if least_observed < _MIN_SAMPLES_PER_MODEL:
+        return next(model for model in ordered if counts[model] == least_observed)
+    return _best_observed_model(state, candidates, ordered, objective)
+
+
+def _seconds(value: object) -> float | None:
+    if isinstance(value, datetime):
+        return value.timestamp()
+    if isinstance(value, timedelta):
+        return value.total_seconds()
+    if isinstance(value, bool) or not isinstance(value, int | float):
+        return None
+    return float(value)
+
+
+def _field(value: object, key: str) -> object | None:
+    if isinstance(value, Mapping):
+        try:
+            return _OBJECT_MAPPING_ADAPTER.validate_python(value).get(key)
+        except ValidationError:
+            return None
+    try:
+        return _OBJECT_ADAPTER.validate_python(object.__getattribute__(value, key))
+    except AttributeError:
+        return None
+
+
+def _visible_output_tokens(response_obj: object) -> int | None:
+    usage = _field(response_obj, "usage")
+    if usage is None:
+        return None
+    total = _field(usage, "completion_tokens")
+    if total is None:
+        total = _field(usage, "output_tokens")
+    if isinstance(total, bool) or not isinstance(total, int | float) or total <= 0:
+        return None
+    details = _field(usage, "completion_tokens_details")
+    if details is None:
+        details = _field(usage, "output_tokens_details")
+    reasoning = _field(details, "reasoning_tokens") if details is not None else None
+    reasoning_tokens = float(reasoning) if isinstance(reasoning, int | float) and not isinstance(reasoning, bool) else 0
+    return max(1, round(float(total) - reasoning_tokens))
+
+
+def response_latency_sample(
+    kwargs: Mapping[str, object], response_obj: object, start_time: object, end_time: object
+) -> tuple[ResponseLatencySample, int] | None:
+    start: Final = _seconds(start_time)
+    end: Final = _seconds(end_time)
+    visible_tokens: Final = _visible_output_tokens(response_obj)
+    if start is None or end is None or end <= start or visible_tokens is None:
+        return None
+    total_ms: Final = (end - start) * 1000
+    completion_start: Final = _seconds(kwargs.get("completion_start_time"))
+    ttft_ms = (
+        (completion_start - start) * 1000 if completion_start is not None and start < completion_start < end else 0.0
+    )
+    generation_ms: Final = max(total_ms - ttft_ms, 1e-6)
+    return ResponseLatencySample(
+        ttft_ms=ttft_ms,
+        output_tokens_per_ms=visible_tokens / generation_ms,
+    ), visible_tokens
+
+
+async def record_runtime_response_latency(
+    *,
+    router_cache: ResponseLatencyCache,
+    router_model_name: str,
+    tier: str,
+    routed_model: str,
+    kwargs: Mapping[str, object],
+    response_obj: object,
+    start_time: object,
+    end_time: object,
+) -> None:
+    measured: Final = response_latency_sample(kwargs, response_obj, start_time, end_time)
+    if measured is None:
+        return
+    sample, visible_tokens = measured
+    key: Final = _cache_key(router_model_name, tier)
+    state: Final = _state(await router_cache.async_get_cache(key=key))
+    samples: Final = (*state.samples_by_model.get(routed_model, ()), sample)[-_MAX_SAMPLES_PER_MODEL:]
+    updated_state: Final = ResponseLatencyState(
+        samples_by_model=MappingProxyType({**state.samples_by_model, routed_model: samples}),
+        visible_output_tokens=(*state.visible_output_tokens, visible_tokens)[-_MAX_EXPECTED_OUTPUT_SAMPLES:],
+    )
+    await router_cache.async_set_cache(key=key, value=updated_state.model_dump(mode="json"), ttl=_CACHE_TTL_SECONDS)
diff --git a/litellm/types/management_endpoints/auto_router_endpoints.py b/litellm/types/management_endpoints/auto_router_endpoints.py
index adbd4486032..7405a3720d6 100644
--- a/litellm/types/management_endpoints/auto_router_endpoints.py
+++ b/litellm/types/management_endpoints/auto_router_endpoints.py
@@ -9,8 +9,11 @@ from typing import Final, Literal, TypeAlias
 
 from pydantic import BaseModel, Field, computed_field, field_validator, model_validator
 
-from litellm.router_strategy.complexity_router.auto_setup import AutoSetupQualityLevel
-from litellm.router_strategy.complexity_router.config import ComplexityRouterConfig
+from litellm.router_strategy.complexity_router.config import (
+    AutoSetupObjective,
+    AutoSetupQualityLevel,
+    ComplexityRouterConfig,
+)
 from litellm.types.utils import StandardLoggingRoutingDecision
 
 DEFAULT_ROUTING_TEST_ROUTER_NAME: Final[str] = "auto_router_routing_test"
@@ -49,13 +52,14 @@ class AutoRouterExcludedModelGroup(BaseModel):
     """A caller-visible reason Auto setup did not use an available model group."""
 
     model_group: str
-    reason: Literal["no_benchmark_match", "mixed_model_group"]
+    reason: Literal["no_benchmark_match", "mixed_model_group", "pricing_unavailable"]
 
 
 class AutoRouterRecommendationResponse(BaseModel):
     """An editable complexity-router config generated from the caller's usable model groups."""
 
     quality_level: AutoSetupQualityLevel
+    optimize_for: AutoSetupObjective
     snapshot_id: str
     snapshot_generated_at: str
     available_model_group_count: int = Field(ge=0)
diff --git a/litellm/types/utils.py b/litellm/types/utils.py
index 34c47d3f201..74a682b8f15 100644
--- a/litellm/types/utils.py
+++ b/litellm/types/utils.py
@@ -2938,6 +2938,16 @@ class StandardLoggingRoutingDecision(TypedDict, total=False):
     savings_baseline_model: str
     savings_baseline_deployment_id: str
     tier_litellm_params: Mapping[str, object]  # writable-ok: Pydantic warns on ReadOnly TypedDict fields
+    auto_setup_snapshot_id: str  # writable-ok: Pydantic warns on ReadOnly TypedDict fields
+    auto_setup_quality_level: Literal[  # writable-ok: Pydantic rejects ReadOnly fields
+        "economy", "balanced", "high", "max"
+    ]
+    auto_setup_optimize_for: Literal[  # writable-ok: Pydantic rejects ReadOnly fields
+        "cost", "task_completion_speed", "balanced"
+    ]
+    auto_setup_selection_mode: Literal[  # writable-ok: Pydantic rejects ReadOnly fields
+        "snapshot_ranked", "runtime_response_latency"
+    ]
 
 
 # Fields whose values quote the caller's prompt. Dropped when an operator turns message
@@ -2966,6 +2976,10 @@ DERIVED_ROUTING_DECISION_FIELDS: Final[frozenset[str]] = frozenset(
         "savings_baseline_model",
         "savings_baseline_deployment_id",
         "tier_litellm_params",
+        "auto_setup_snapshot_id",
+        "auto_setup_quality_level",
+        "auto_setup_optimize_for",
+        "auto_setup_selection_mode",
     }
 )
 
diff --git a/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py
index 278fb50e1dd..ccaf4d5a0e9 100644
--- a/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py
+++ b/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py
@@ -136,6 +136,7 @@ async def test_recommendation_uses_caller_groups_and_returns_an_editable_complex
     response = await get_auto_router_recommendation(
         user_api_key_dict=ADMIN,
         quality_level="economy",
+        optimize_for="task_completion_speed",
     )
 
     authorize.assert_awaited_once_with(user_api_key_dict=ADMIN, team_id=None)
@@ -144,8 +145,11 @@ async def test_recommendation_uses_caller_groups_and_returns_an_editable_complex
     assert response.excluded_model_groups == ()
     assert "unavailable-group" not in response.matched_model_groups
     assert response.complexity_router_config.classifier_type == "heuristic_v2"
-    assert all(len(models) == 1 for models in response.complexity_router_config.tiers.values())
-    assert "auto_setup" not in response.complexity_router_config.model_dump(mode="json", exclude_none=True)
+    assert response.complexity_router_config.auto_setup is not None
+    assert response.complexity_router_config.auto_setup.optimize_for == "task_completion_speed"
+    assert response.complexity_router_config.auto_setup.tier_policies["SIMPLE"].selection_mode == (
+        "runtime_response_latency"
+    )
 
 
 def test_auto_setup_pricing_honors_custom_token_rates_and_rejects_non_token_charges() -> None:
@@ -194,7 +198,7 @@ def test_auto_setup_pricing_resolves_an_explicit_base_model_before_a_deployment_
 
 
 @pytest.mark.asyncio
-async def test_recommendation_excludes_mixed_and_unmatched_but_keeps_unpriced_groups(
+async def test_recommendation_excludes_mixed_unmatched_and_unpriced_groups(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
     import litellm.proxy.management_endpoints.auto_router_endpoints as endpoint_module
@@ -222,14 +226,15 @@ async def test_recommendation_excludes_mixed_and_unmatched_but_keeps_unpriced_gr
     response = await get_auto_router_recommendation(
         user_api_key_dict=ADMIN,
         quality_level="max",
+        optimize_for="cost",
     )
 
     assert response.available_model_group_count == 4
-    assert set(response.matched_model_groups) <= {"safe", "unpriced"}
-    assert response.matched_model_groups
+    assert response.matched_model_groups == ("safe",)
     assert {(item.model_group, item.reason) for item in response.excluded_model_groups} == {
         ("mixed", "mixed_model_group"),
         ("unknown", "no_benchmark_match"),
+        ("unpriced", "pricing_unavailable"),
     }
 
 
diff --git a/tests/test_litellm/router_strategy/test_complexity_router_auto_setup.py b/tests/test_litellm/router_strategy/test_complexity_router_auto_setup.py
index 64a2e46f09a..9a19a5fb12c 100644
--- a/tests/test_litellm/router_strategy/test_complexity_router_auto_setup.py
+++ b/tests/test_litellm/router_strategy/test_complexity_router_auto_setup.py
@@ -1,21 +1,32 @@
 import math
 import random
-from unittest.mock import MagicMock, patch
+from itertools import product
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 from pydantic import ValidationError
 
+from litellm import Router
 from litellm.router_strategy.complexity_router.auto_setup import (
     AutoRouterSnapshot,
     AutoSetupDeployment,
     AutoSetupDeploymentPricing,
-    _quality_completion_cost,  # pyright: ignore[reportPrivateUsage] -- regression test for deployment-aware setup ranking
+    _speed_completion_cost,  # pyright: ignore[reportPrivateUsage] -- regression test for multi-call trajectory pricing
     analyze_auto_setup_inventory,
     build_auto_setup_config,
     load_auto_router_snapshot,
 )
 from litellm.router_strategy.complexity_router.complexity_router import ComplexityRouter
-from litellm.router_strategy.complexity_router.config import ComplexityTier
+from litellm.router_strategy.complexity_router.config import (
+    AutoSetupCandidate,
+    ComplexityRouterConfig,
+    ComplexityTier,
+)
+from litellm.router_strategy.complexity_router.response_latency import (
+    record_runtime_response_latency,
+    response_latency_sample,
+    select_runtime_response_model,
+)
 
 
 def _available_refs(*model_names: str) -> dict[str, tuple[AutoSetupDeployment, ...]]:
@@ -51,38 +62,44 @@ def test_snapshot_is_verified_and_quality_gate_is_recomputed_for_available_model
         snapshot=snapshot,
         available_model_deployments=available,
         quality_level="economy",
+        optimize_for="cost",
     )
     maximum = build_auto_setup_config(
         snapshot=snapshot,
         available_model_deployments=available,
         quality_level="max",
+        optimize_for="cost",
     )
     nano_only = build_auto_setup_config(
         snapshot=snapshot,
         available_model_deployments=_available_refs("gpt-5.4-nano"),
         quality_level="max",
+        optimize_for="cost",
     )
 
-    assert snapshot.artifact_sha256 == "e191b5f7225e4da7a178eb5fe8780bcea24fb0166be23e2069338c6fe5522cd5"
-    assert economy.tiers["SIMPLE"] == ["gpt-5.4-nano"]
+    assert snapshot.artifact_sha256 == "bab7488212d40d96c5e43f80893c5ca2961d0e25160cb1aebbef36a2422b8a2b"
+    assert economy.tiers["SIMPLE"] == ["gpt-5.4-nano", "gpt-5.6-sol"]
     assert maximum.tiers["SIMPLE"] == ["gpt-5.6-sol"]
     assert nano_only.tiers["SIMPLE"] == ["gpt-5.4-nano"]
 
 
-def test_auto_setup_emits_a_plain_static_complexity_router_config() -> None:
+def test_speed_only_changes_auto_generated_selection_policy() -> None:
     config = build_auto_setup_config(
         snapshot=load_auto_router_snapshot(),
         available_model_deployments=_available_refs("gpt-5.6-sol", "gpt-5.4-nano"),
         quality_level="economy",
+        optimize_for="task_completion_speed",
     )
 
-    payload = config.model_dump(mode="json", exclude_none=True)
-    assert payload["classifier_type"] == "heuristic_v2"
-    assert "auto_setup" not in payload
-    assert all(len(models) == 1 for models in config.tiers.values())
+    assert config.auto_setup is not None
+    assert config.auto_setup.tier_policies["SIMPLE"].selection_mode == "runtime_response_latency"
+    assert config.auto_setup.tier_policies["MEDIUM"].selection_mode == "runtime_response_latency"
+    assert config.auto_setup.tier_policies["COMPLEX"].selection_mode == "snapshot_ranked"
+    assert config.auto_setup.tier_policies["REASONING"].selection_mode == "snapshot_ranked"
+    assert ComplexityRouterConfig(tiers=config.tiers).auto_setup is None
 
 
-def test_every_snapshot_alias_builds_all_four_quality_setups_even_without_price() -> None:
+def test_every_priced_snapshot_alias_builds_all_twelve_auto_setups() -> None:
     snapshot = load_auto_router_snapshot()
     aliases = tuple(
         alias
@@ -101,17 +118,25 @@ def test_every_snapshot_alias_builds_all_four_quality_setups_even_without_price(
     )
 
     for alias, deployment in priced.items():
-        eligible, excluded = analyze_auto_setup_inventory(snapshot, {alias: (deployment,)})
-        assert eligible == (alias,)
-        assert excluded == ()
-        for quality_level in ("economy", "balanced", "high", "max"):
+        if deployment.pricing is None:
+            eligible, excluded = analyze_auto_setup_inventory(snapshot, {alias: (deployment,)})
+            assert not eligible
+            assert [(item.model_group, item.reason) for item in excluded] == [(alias, "pricing_unavailable")]
+            continue
+        for quality_level, objective in product(
+            ("economy", "balanced", "high", "max"),
+            ("cost", "task_completion_speed", "balanced"),
+        ):
             config = build_auto_setup_config(
                 snapshot=snapshot,
                 available_model_deployments={alias: (deployment,)},
                 quality_level=quality_level,
+                optimize_for=objective,
             )
             assert set(config.tiers) == {"SIMPLE", "MEDIUM", "COMPLEX", "REASONING"}
             assert all(models == [alias] for models in config.tiers.values())
+            assert config.auto_setup is not None
+            assert all(policy.candidates[0].model_name == alias for policy in config.auto_setup.tier_policies.values())
 
 
 def test_full_priced_inventory_builds_deterministically_for_every_choice() -> None:
@@ -130,24 +155,33 @@ def test_full_priced_inventory_builds_deterministically_for_every_choice() -> No
     }
     assert len(deployments) == 279
 
-    selected_tiers: dict[str, dict[str, set[str]]] = {}
-    for quality_level in ("economy", "balanced", "high", "max"):
+    cost_tiers: dict[str, dict[str, set[str]]] = {}
+    for quality_level, objective in product(
+        ("economy", "balanced", "high", "max"),
+        ("cost", "task_completion_speed", "balanced"),
+    ):
         config = build_auto_setup_config(
             snapshot=snapshot,
             available_model_deployments=deployments,
             quality_level=quality_level,
+            optimize_for=objective,
         )
         repeated = build_auto_setup_config(
             snapshot=snapshot,
             available_model_deployments=dict(reversed(tuple(deployments.items()))),
             quality_level=quality_level,
+            optimize_for=objective,
         )
         assert config.model_dump(mode="json") == repeated.model_dump(mode="json")
-        assert all(len(models) == 1 for models in config.tiers.values())
+        assert all(models and len(models) == len(set(models)) for models in config.tiers.values())
         assert all(set(models).issubset(deployments) for models in config.tiers.values())
-        selected_tiers[quality_level] = {tier: set(models) for tier, models in config.tiers.items()}
+        if objective == "cost":
+            cost_tiers[quality_level] = {tier: set(models) for tier, models in config.tiers.items()}
 
-    assert set(selected_tiers) == {"economy", "balanced", "high", "max"}
+    for tier in ("SIMPLE", "MEDIUM", "COMPLEX", "REASONING"):
+        assert cost_tiers["max"][tier] <= cost_tiers["high"][tier]
+        assert cost_tiers["high"][tier] <= cost_tiers["balanced"][tier]
+        assert cost_tiers["balanced"][tier] <= cost_tiers["economy"][tier]
 
 
 def test_seeded_mixed_inventories_never_select_unknown_or_unsafe_groups() -> None:
@@ -179,11 +213,15 @@ def test_seeded_mixed_inventories_never_select_unknown_or_unsafe_groups() -> Non
                 for index in range(rng.randint(0, 10))
             }
         )
-        for quality_level in ("economy", "balanced", "high", "max"):
+        for quality_level, objective in product(
+            ("economy", "balanced", "high", "max"),
+            ("cost", "task_completion_speed", "balanced"),
+        ):
             config = build_auto_setup_config(
                 snapshot=snapshot,
                 available_model_deployments=inventory,
                 quality_level=quality_level,
+                optimize_for=objective,
             )
             selected = {model_name for models in config.tiers.values() for model_name in models}
             assert selected
@@ -220,11 +258,12 @@ def test_mixed_or_partially_unmatched_model_groups_fail_closed() -> None:
         snapshot=snapshot,
         available_model_deployments=inventory,
         quality_level="economy",
+        optimize_for="cost",
     )
     assert all(models == ["safe"] for models in config.tiers.values())
 
 
-def test_missing_cost_falls_back_to_quality_instead_of_blocking_setup() -> None:
+def test_schema_valid_missing_cost_and_zero_completion_evidence_fail_closed() -> None:
     raw = load_auto_router_snapshot().model_dump(mode="json")
     quality_profiles = raw["models"]["gpt-5.4-nano-xhigh"]["quality_by_complexity"].values()
     for profile in quality_profiles:
@@ -235,29 +274,8 @@ def test_missing_cost_falls_back_to_quality_instead_of_blocking_setup() -> None:
     snapshot = AutoRouterSnapshot.model_validate(raw)
     eligible, excluded = analyze_auto_setup_inventory(snapshot, _available_refs("gpt-5.4-nano"))
 
-    assert eligible == ("gpt-5.4-nano",)
-    assert excluded == ()
-    generated = build_auto_setup_config(
-        snapshot=snapshot,
-        available_model_deployments=_available_refs("gpt-5.4-nano"),
-        quality_level="max",
-    )
-    assert all(models == ["gpt-5.4-nano"] for models in generated.tiers.values())
-
-
-def test_multiple_unpriced_models_choose_the_highest_quality_inside_the_gate() -> None:
-    inventory = {
-        model_name: (AutoSetupDeployment(model_refs=(model_name,), pricing=None),)
-        for model_name in ("gpt-5.6-sol", "gpt-5.4-nano")
-    }
-
-    generated = build_auto_setup_config(
-        snapshot=load_auto_router_snapshot(),
-        available_model_deployments=inventory,
-        quality_level="max",
-    )
-
-    assert generated.tiers["SIMPLE"] == ["gpt-5.6-sol"]
+    assert eligible == ()
+    assert [(item.model_group, item.reason) for item in excluded] == [("gpt-5.4-nano", "pricing_unavailable")]
 
 
 def test_provider_repricing_and_multi_deployment_groups_are_conservative() -> None:
@@ -270,23 +288,42 @@ def test_provider_repricing_and_multi_deployment_groups_are_conservative() -> No
         model_refs=("gpt-5.6-sol",),
         pricing=AutoSetupDeploymentPricing(input_cost_per_token=0.00001, output_cost_per_token=0.00002),
     )
-    profile = snapshot.models["gpt-5.6-sol-max"].quality_by_complexity["trivial"]
-    cheap_cost = _quality_completion_cost(profile, (cheap,))
-    expensive_cost = _quality_completion_cost(profile, (expensive,))
-    mixed_cost = _quality_completion_cost(profile, (cheap, expensive))
-
-    assert cheap_cost is not None
-    assert expensive_cost is not None
-    assert mixed_cost is not None
-    assert math.isclose(expensive_cost, cheap_cost * 10)
-    assert math.isclose(mixed_cost, expensive_cost)
-
     config = build_auto_setup_config(
         snapshot=snapshot,
         available_model_deployments={"cheap": (cheap,), "expensive": (expensive,), "mixed-price": (cheap, expensive)},
         quality_level="economy",
+        optimize_for="cost",
     )
-    assert config.tiers["SIMPLE"] == ["cheap"]
+    assert config.auto_setup is not None
+    candidates = config.auto_setup.tier_policies["SIMPLE"].candidates
+    by_name = {candidate.model_name: candidate.cost_per_completed_task_usd for candidate in candidates}
+    assert tuple(candidate.model_name for candidate in candidates) == ("cheap", "expensive", "mixed-price")
+    assert math.isclose(by_name["expensive"], by_name["cheap"] * 10)
+    assert math.isclose(by_name["mixed-price"], by_name["expensive"])
+
+
+def test_multi_call_trajectory_cost_does_not_apply_single_request_context_thresholds() -> None:
+    snapshot = load_auto_router_snapshot()
+    speed = snapshot.models["gpt-5.6-sol-max"].task_completion_speed_by_complexity["complex"][0]
+    pricing = AutoSetupDeploymentPricing(
+        input_cost_per_token=0.000001,
+        output_cost_per_token=0.000002,
+        cache_read_input_token_cost=0.0000001,
+        input_cost_per_token_above_200k_tokens=0.001,
+        output_cost_per_token_above_200k_tokens=0.002,
+        cache_read_input_token_cost_above_200k_tokens=0.0001,
+    )
+    deployment = AutoSetupDeployment(model_refs=("gpt-5.6-sol",), pricing=pricing)
+
+    actual = _speed_completion_cost(speed, (deployment,))
+    expected_attempt_cost = (
+        (speed.mean_uncached_input_tokens or 0) * pricing.input_cost_per_token
+        + (speed.mean_cache_read_input_tokens or 0) * (pricing.cache_read_input_token_cost or 0)
+        + (speed.mean_output_tokens or 0) * pricing.output_cost_per_token
+    )
+
+    assert actual is not None
+    assert math.isclose(actual, expected_attempt_cost / speed.success_probability)
 
 
 @pytest.mark.parametrize(
@@ -315,27 +352,164 @@ def test_snapshot_semantic_validation_rejects_inconsistent_artifacts(mutation: s
         AutoRouterSnapshot.model_validate(raw)
 
 
-def test_auto_setup_result_has_no_runtime_only_configuration() -> None:
+def test_auto_setup_cannot_change_adaptive_or_custom_tier_routers() -> None:
     generated = build_auto_setup_config(
         snapshot=load_auto_router_snapshot(),
         available_model_deployments=_available_refs("gpt-5.6-sol"),
         quality_level="max",
+        optimize_for="cost",
     )
-    payload = generated.model_dump(mode="json", exclude_none=True)
+    payload = generated.model_dump(mode="json")
 
-    assert "auto_setup" not in payload
-    assert payload["adaptive"] is False
-    assert payload["classifier_type"] == "heuristic_v2"
+    with pytest.raises(ValidationError, match="auto_setup and adaptive"):
+        ComplexityRouterConfig.model_validate({**payload, "adaptive": True})
+    with pytest.raises(ValidationError, match="built-in four-tier"):
+        ComplexityRouterConfig.model_validate(
+            {
+                **payload,
+                "tier_definitions": [
+                    {"name": "fast", "description": "Fast"},
+                    {"name": "strong", "description": "Strong"},
+                ],
+                "tiers": {"fast": ["gpt-5.6-sol"], "strong": ["gpt-5.6-sol"]},
+                "fallback_tier": "fast",
+                "classifier_type": "llm",
+                "classifier_llm_config": {"model": "gpt-5.6-sol"},
+            }
+        )
+
+
+class _Cache:
+    def __init__(self) -> None:
+        self.values: dict[str, object] = {}
+
+    async def async_get_cache(
+        self,
+        key: str,
+        parent_otel_span: object | None = None,
+        local_only: bool = False,
+        **kwargs: object,
+    ) -> object:
+        return self.values.get(key)
+
+    async def async_set_cache(
+        self,
+        key: str,
+        value: object,
+        local_only: bool = False,
+        **kwargs: object,
+    ) -> None:
+        self.values[key] = value
+
+
+def _candidate(model_name: str, cost: float) -> AutoSetupCandidate:
+    return AutoSetupCandidate(
+        model_name=model_name,
+        benchmark_model_id=f"benchmark-{model_name}",
+        quality_lower_bound=0.95,
+        cost_per_completed_task_usd=cost,
+    )
 
 
 @pytest.mark.asyncio
-async def test_generated_single_model_tiers_use_the_existing_router_selection_path() -> None:
+async def test_easy_task_speed_learns_equal_output_response_time() -> None:
+    cache = _Cache()
+    candidates = (_candidate("cheap-slow", 0.01), _candidate("fast", 0.04))
+
+    assert (
+        await select_runtime_response_model(
+            router_cache=cache,
+            router_model_name="auto",
+            tier="SIMPLE",
+            candidates=candidates,
+            available_models=("cheap-slow", "fast"),
+            cold_start_model="cheap-slow",
+            objective="task_completion_speed",
+        )
+        == "cheap-slow"
+    )
+
+    for model, end in (("cheap-slow", 0.30), ("cheap-slow", 0.30), ("fast", 0.15), ("fast", 0.15)):
+        await record_runtime_response_latency(
+            router_cache=cache,
+            router_model_name="auto",
+            tier="SIMPLE",
+            routed_model=model,
+            kwargs={"completion_start_time": 0.10},
+            response_obj={
+                "usage": {
+                    "completion_tokens": 20,
+                    "completion_tokens_details": {"reasoning_tokens": 10},
+                }
+            },
+            start_time=0.0,
+            end_time=end,
+        )
+
+    measured = response_latency_sample(
+        {"completion_start_time": 0.10},
+        {"usage": {"completion_tokens": 20, "completion_tokens_details": {"reasoning_tokens": 10}}},
+        0.0,
+        0.30,
+    )
+    assert measured is not None and measured[1] == 10
+    assert (
+        await select_runtime_response_model(
+            router_cache=cache,
+            router_model_name="auto",
+            tier="SIMPLE",
+            candidates=candidates,
+            available_models=("cheap-slow", "fast"),
+            cold_start_model="cheap-slow",
+            objective="task_completion_speed",
+        )
+        == "fast"
+    )
+
+
+@pytest.mark.asyncio
+async def test_easy_task_balanced_uses_both_live_response_time_and_expected_cost() -> None:
+    cache = _Cache()
+    candidates = (
+        _candidate("fast-expensive", 1.0),
+        _candidate("middle", 0.02),
+        _candidate("cheap-slow", 0.01),
+    )
+    for model, end in (("fast-expensive", 0.10), ("middle", 0.12), ("cheap-slow", 0.30)):
+        for _ in range(2):
+            await record_runtime_response_latency(
+                router_cache=cache,
+                router_model_name="auto",
+                tier="SIMPLE",
+                routed_model=model,
+                kwargs={},
+                response_obj={"usage": {"completion_tokens": 10}},
+                start_time=0.0,
+                end_time=end,
+            )
+
+    common = {
+        "router_cache": cache,
+        "router_model_name": "auto",
+        "tier": "SIMPLE",
+        "candidates": candidates,
+        "available_models": tuple(candidate.model_name for candidate in candidates),
+        "cold_start_model": "cheap-slow",
+    }
+    assert await select_runtime_response_model(**common, objective="task_completion_speed") == "fast-expensive"
+    assert await select_runtime_response_model(**common, objective="balanced") == "middle"
+
+
+@pytest.mark.asyncio
+async def test_snapshot_policy_is_deterministic_but_manual_pool_keeps_random_selection() -> None:
     generated = build_auto_setup_config(
         snapshot=load_auto_router_snapshot(),
         available_model_deployments=_available_refs("gpt-5.6-sol", "gpt-5.4-nano"),
         quality_level="economy",
+        optimize_for="cost",
     )
     router_instance = MagicMock()
+    router_instance.cache = AsyncMock()
     automatic = ComplexityRouter(
         model_name="auto",
         litellm_router_instance=router_instance,
@@ -348,8 +522,71 @@ async def test_generated_single_model_tiers_use_the_existing_router_selection_pa
     )
 
     assert await automatic._pick_model_for_tier(ComplexityTier.SIMPLE, None, None, {}) == "gpt-5.4-nano"
+    assert (
+        await automatic._pick_model_for_tier(
+            ComplexityTier.SIMPLE,
+            None,
+            None,
+            {},
+            allowed_models=("gpt-5.6-sol",),
+        )
+        == "gpt-5.6-sol"
+    )
     with patch(  # test-quality-ok: pinning random choice is required to prove manual routers retain their existing selection path
         "litellm.router_strategy.complexity_router.complexity_router.random.choice", return_value="second"
     ) as pick:
         assert await manual._pick_model_for_tier(ComplexityTier.SIMPLE, None, None, {}) == "second"
-    pick.assert_called_once_with(["first", "second"])
+    pick.assert_called_once_with(("first", "second"))
+
+
+@pytest.mark.asyncio
+async def test_live_latency_telemetry_is_recorded_only_for_an_auto_runtime_policy() -> None:
+    # test-quality-ok: this callback's observable contract is whether telemetry is emitted
+    generated = build_auto_setup_config(
+        snapshot=load_auto_router_snapshot(),
+        available_model_deployments=_available_refs("gpt-5.6-sol", "gpt-5.4-nano"),
+        quality_level="economy",
+        optimize_for="task_completion_speed",
+    )
+    router = Router(
+        model_list=[
+            {
+                "model_name": "smart",
+                "litellm_params": {
+                    "model": "auto_router/complexity_router",
+                    "complexity_router_config": generated.model_dump(mode="json"),
+                    "complexity_router_default_model": "gpt-5.6-sol",
+                },
+            },
+            {"model_name": "gpt-5.6-sol", "litellm_params": {"model": "gpt-5.6-sol"}},
+            {"model_name": "gpt-5.4-nano", "litellm_params": {"model": "gpt-5.4-nano"}},
+        ]
+    )
+    metadata = {
+        "routing_decision": {
+            "router_model_name": "smart",
+            "tier": "SIMPLE",
+            "routed_model": "gpt-5.4-nano",
+            "auto_setup_selection_mode": "runtime_response_latency",
+        }
+    }
+
+    with patch(  # test-quality-ok: patching the telemetry boundary avoids mutating the shared runtime latency cache
+        "litellm.router_strategy.complexity_router.response_latency.record_runtime_response_latency",
+        new_callable=AsyncMock,
+    ) as recorder:
+        await router._record_auto_setup_response_latency(
+            kwargs={"litellm_params": {"metadata": metadata}},
+            completion_response={"usage": {"completion_tokens": 1}},
+            start_time=0.0,
+            end_time=0.1,
+        )
+        metadata["routing_decision"]["auto_setup_selection_mode"] = "snapshot_ranked"
+        await router._record_auto_setup_response_latency(
+            kwargs={"litellm_params": {"metadata": metadata}},
+            completion_response={"usage": {"completion_tokens": 1}},
+            start_time=0.0,
+            end_time=0.1,
+        )
+
+    recorder.assert_awaited_once()
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index e4236afc586..35f4f1c1a14 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -12703,7 +12703,7 @@ async def test_prompt_management_factory_marks_injection_for_every_deployment(mo
     assert bucket["litellm_gateway_injected_cache"] == ""
 
 
-def test_get_configured_mode_reads_deployment_model_info():
+def test_get_configured_mode_reads_audio_speech_deployment_model_info():
     router = Router(
         model_list=[
             {
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/autoRouter/useAutoRouterRecommendation.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/autoRouter/useAutoRouterRecommendation.ts
index 8a5b35e22ce..d13452ffd9a 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/autoRouter/useAutoRouterRecommendation.ts
+++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/autoRouter/useAutoRouterRecommendation.ts
@@ -1,7 +1,7 @@
 import { useQuery } from "@tanstack/react-query";
 
 import { getAutoRouterRecommendation, type AutoRouterRecommendationResponse } from "@/components/networking";
-import type { AutoSetupQualityLevel } from "@/components/add_model/build_complexity_router_config";
+import type { AutoSetupObjective, AutoSetupQualityLevel } from "@/components/add_model/build_complexity_router_config";
 import { createQueryKeys } from "../common/queryKeysFactory";
 
 const recommendationKeys = createQueryKeys("autoRouterRecommendation");
@@ -9,6 +9,7 @@ const recommendationKeys = createQueryKeys("autoRouterRecommendation");
 interface UseAutoRouterRecommendationParams {
   accessToken: string;
   qualityLevel: AutoSetupQualityLevel;
+  optimizeFor: AutoSetupObjective;
   teamId?: string;
   enabled: boolean;
 }
@@ -16,15 +17,16 @@ interface UseAutoRouterRecommendationParams {
 export const useAutoRouterRecommendation = ({
   accessToken,
   qualityLevel,
+  optimizeFor,
   teamId,
   enabled,
 }: UseAutoRouterRecommendationParams) => {
-  const filters = { qualityLevel, ...(teamId ? { teamId } : {}) };
+  const filters = { qualityLevel, optimizeFor, ...(teamId ? { teamId } : {}) };
   const queryOptions = {
     queryKey: recommendationKeys.list({
       filters,
     }),
-    queryFn: () => getAutoRouterRecommendation(accessToken, qualityLevel, teamId),
+    queryFn: () => getAutoRouterRecommendation(accessToken, qualityLevel, optimizeFor, teamId),
     enabled: enabled && Boolean(accessToken),
     staleTime: 30 * 1000,
   };
diff --git a/ui/litellm-dashboard/src/components/add_model/AutoSetupControls.tsx b/ui/litellm-dashboard/src/components/add_model/AutoSetupControls.tsx
index e1252b2375b..d506c43af3a 100644
--- a/ui/litellm-dashboard/src/components/add_model/AutoSetupControls.tsx
+++ b/ui/litellm-dashboard/src/components/add_model/AutoSetupControls.tsx
@@ -2,7 +2,13 @@ import type { AutoRouterRecommendationResponse } from "@/components/networking";
 import { Button } from "@/components/ui/button";
 import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select";
 import { UiLoadingSpinner } from "@/components/ui/ui-loading-spinner";
-import type { AutoSetupQualityLevel } from "./build_complexity_router_config";
+import type { AutoSetupObjective, AutoSetupQualityLevel } from "./build_complexity_router_config";
+
+const OBJECTIVE_DESCRIPTION: Record<AutoSetupObjective, string> = {
+  cost: "Spend less while staying within your selected quality level",
+  task_completion_speed: "Finish requests sooner, not just generate tokens faster",
+  balanced: "Balance expected cost with time to completion",
+};
 
 const QUALITY_OPTIONS = [
   { value: "economy", label: "Economy" },
@@ -11,23 +17,26 @@ const QUALITY_OPTIONS = [
   { value: "max", label: "Max" },
 ];
 
-const QUALITY_DESCRIPTION: Record<AutoSetupQualityLevel, string> = {
-  economy: "Allows models up to 15 points below the smartest available option",
-  balanced: "Keeps models within 7 points of the smartest available option",
-  high: "Keeps models within 3 points of the smartest available option",
-  max: "Keeps models within 1 point of the smartest available option",
-};
+const OBJECTIVE_OPTIONS = [
+  { value: "cost", label: "Cost" },
+  { value: "task_completion_speed", label: "Task completion speed" },
+  { value: "balanced", label: "Balanced" },
+];
 
 interface AutoSetupControlsProps {
   qualityLevel: AutoSetupQualityLevel;
+  optimizeFor: AutoSetupObjective;
   onQualityLevelChange: (value: AutoSetupQualityLevel) => void;
+  onOptimizeForChange: (value: AutoSetupObjective) => void;
   recommendation?: AutoRouterRecommendationResponse;
   recommendationEnabled: boolean;
   recommendationPending: boolean;
   recommendationFailed: boolean;
   recommendationError: unknown;
   requiresTeamScope: boolean;
+  hasAutoPolicy: boolean;
   onRetry: () => void;
+  onRegenerate: () => void;
   onBack: () => void;
 }
 
@@ -40,16 +49,31 @@ const readyMessage = (recommendation: AutoRouterRecommendationResponse): string
   return `Uses ${count} of ${total} available model${total === 1 ? "" : "s"} across four complexity levels`;
 };
 
+const recommendationDescription = (optimizeFor: AutoSetupObjective): string => {
+  switch (optimizeFor) {
+    case "cost":
+      return "Favors the lowest expected cost while keeping each request within your selected quality level";
+    case "task_completion_speed":
+      return "Easy requests use live response speed; difficult requests use benchmark completion time";
+    case "balanced":
+      return "Balances expected cost with response speed and benchmark completion time";
+  }
+};
+
 export default function AutoSetupControls({
   qualityLevel,
+  optimizeFor,
   onQualityLevelChange,
+  onOptimizeForChange,
   recommendation,
   recommendationEnabled,
   recommendationPending,
   recommendationFailed,
   recommendationError,
   requiresTeamScope,
+  hasAutoPolicy,
   onRetry,
+  onRegenerate,
   onBack,
 }: AutoSetupControlsProps) {
   const excludedCount = recommendation?.excluded_model_groups.length ?? 0;
@@ -62,7 +86,7 @@ export default function AutoSetupControls({
           <div className="text-xs text-muted-foreground">
             Choose the outcome you want. LiteLLM will build the model tiers for you
           </div>
-          {recommendation && excludedCount > 0 && (
+          {hasAutoPolicy && excludedCount > 0 && (
             <div className="text-xs text-muted-foreground">
               {excludedCount} model{excludedCount === 1 ? " was" : "s were"} left out because Auto setup could not
               compare {excludedCount === 1 ? "it" : "them"} safely
@@ -73,7 +97,7 @@ export default function AutoSetupControls({
           Back
         </Button>
       </div>
-      <div>
+      <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
         <div>
           <label className="mb-2 block text-sm font-medium text-foreground">Quality level</label>
           <Select
@@ -92,13 +116,27 @@ export default function AutoSetupControls({
             </SelectContent>
           </Select>
           <div className="mt-1 text-xs text-muted-foreground">
-            {QUALITY_DESCRIPTION[qualityLevel]}. Auto picks the lowest estimated completion cost inside that range, or
-            the smartest option when price is unavailable
-          </div>
-          <div className="mt-1 text-xs text-muted-foreground">
-            Quality uses uncertainty-adjusted public benchmark scores for each complexity level
+            Higher levels keep each tier closer to the smartest available model
           </div>
         </div>
+        <div>
+          <label className="mb-2 block text-sm font-medium text-foreground">Optimize for</label>
+          <Select
+            items={OBJECTIVE_OPTIONS}
+            value={optimizeFor}
+            onValueChange={(value: AutoSetupObjective | null) => value && onOptimizeForChange(value)}
+          >
+            <SelectTrigger data-testid="auto-objective-selector" className="w-full">
+              <SelectValue />
+            </SelectTrigger>
+            <SelectContent>
+              <SelectItem value="cost">Cost</SelectItem>
+              <SelectItem value="task_completion_speed">Task completion speed</SelectItem>
+              <SelectItem value="balanced">Balanced</SelectItem>
+            </SelectContent>
+          </Select>
+          <div className="mt-1 text-xs text-muted-foreground">{OBJECTIVE_DESCRIPTION[optimizeFor]}</div>
+        </div>
       </div>
       {recommendationPending && recommendation === undefined && (
         <div className="flex items-center gap-2 text-xs text-muted-foreground">
@@ -118,11 +156,19 @@ export default function AutoSetupControls({
       )}
       {recommendation && (
         <div className="space-y-2 rounded-lg border border-border bg-muted/40 p-3" data-testid="auto-setup-summary">
-          <div className="text-sm font-medium text-foreground">Recommended setup</div>
-          <div className="text-xs text-muted-foreground">
-            {readyMessage(recommendation)}. Each tier is the lowest-cost model with comparable pricing that meets your
-            selected quality level
+          <div className={`text-sm font-medium ${hasAutoPolicy ? "text-foreground" : "text-warning"}`}>
+            {hasAutoPolicy ? "Recommended setup" : "Configuration edited"}
           </div>
+          <div className="text-xs text-muted-foreground">
+            {hasAutoPolicy
+              ? `${readyMessage(recommendation)}. ${recommendationDescription(optimizeFor)}`
+              : "Your model changes are now treated as a manual configuration"}
+          </div>
+          {!hasAutoPolicy && (
+            <button type="button" className="text-xs text-primary underline" onClick={onRegenerate}>
+              Restore recommended setup
+            </button>
+          )}
         </div>
       )}
     </div>
diff --git a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.test.tsx b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.test.tsx
index 02cb0543e17..62e950c890c 100644
--- a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.test.tsx
+++ b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.test.tsx
@@ -2,6 +2,7 @@ import { fireEvent, renderWithProviders, screen, within } from "../../../tests/t
 import userEvent from "@testing-library/user-event";
 import { vi } from "vitest";
 import ComplexityRouterConfig, { ComplexityRouterConfigValue } from "./ComplexityRouterConfig";
+import { detachAutoSetupAfterPoolEdit } from "./auto_setup_policy";
 vi.mock(
   "@/app/(dashboard)/hooks/autoRouter/useComplexityScorerDefaults",
   async () => await import("../../../tests/mocks/complexityScorerDefaults"),
@@ -44,6 +45,30 @@ const baseProps = {
 };
 
 describe("ComplexityRouterConfig", () => {
+  it("keeps Auto policy for classifier edits and detaches it for model-pool edits", () => {
+    const automatic: ComplexityRouterConfigValue = {
+      ...defaultValue,
+      auto_setup: {
+        schema_version: "1",
+        snapshot_id: "snapshot-test",
+        snapshot_sha256: "a".repeat(64),
+        quality_level: "max",
+        optimize_for: "cost",
+        tier_policies: {},
+      },
+    };
+
+    expect(detachAutoSetupAfterPoolEdit(automatic, { ...automatic, classifier_type: "heuristic_v2" }).auto_setup).toBe(
+      automatic.auto_setup,
+    );
+    expect(
+      detachAutoSetupAfterPoolEdit(automatic, {
+        ...automatic,
+        tiers: { ...automatic.tiers, SIMPLE: ["gpt-4"] },
+      }).auto_setup,
+    ).toBeUndefined();
+  });
+
   it("should render", () => {
     renderWithProviders(<ComplexityRouterConfig {...baseProps} />);
     expect(screen.getByText("Complexity Tier Configuration")).toBeInTheDocument();
diff --git a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx
index 791fd1fcfbb..8e85e215619 100644
--- a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx
+++ b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx
@@ -48,6 +48,7 @@ import EscalationKeywords from "./EscalationKeywords";
 import KeywordTierRules, { KeywordTierRule } from "./KeywordTierRules";
 import SemanticKeywordMatching from "./SemanticKeywordMatching";
 import { type DimensionWeights, type TierBoundaries, type TokenThresholds } from "./heuristic_scoring_knobs";
+import type { AutoSetupConfigPayload } from "./build_complexity_router_config";
 
 export type { DimensionWeights, TierBoundaries, TokenThresholds };
 export type { CustomTierSet, TierRow } from "./tier_rows";
@@ -446,6 +447,8 @@ export interface ComplexityRouterConfigValue {
    * edit round-trip.
    */
   tier_model_params?: TierModelParamsByTier;
+  /** Present only when the zero-setup Auto flow generated this editable config. */
+  auto_setup?: AutoSetupConfigPayload;
 }
 
 /** Session affinity wins where a hand-authored config sets both, matching the backend's own `or`. */
diff --git a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx
index 80be6ed6a81..7f31356c20b 100644
--- a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx
+++ b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx
@@ -4,7 +4,7 @@ import { vi } from "vitest";
 import AddAutoRouterTab from "./add_auto_router_tab";
 import { toast } from "@/lib/toast";
 import { handleAddAutoRouterSubmit } from "./handle_add_auto_router_submit";
-import { getMissingTiersError } from "./build_complexity_router_config";
+import { getMissingTiersError, type AutoSetupConfigPayload } from "./build_complexity_router_config";
 import { getSubmitBlockedReason } from "./add_auto_router_tab";
 import { buildModelAvailability } from "@/lib/autorouter_presets";
 import { testAutoRouterRouting } from "../networking";
@@ -169,17 +169,17 @@ describe("AddAutoRouterTab", () => {
     expect(useAutoRouterRecommendation).toHaveBeenCalledWith(expect.objectContaining({ enabled: false }));
   });
 
-  it("shows the quality choice after Configure automatically is selected", async () => {
+  it("shows the two outcome choices after Configure automatically is selected", async () => {
     const user = userEvent.setup();
     renderWithProviders(<Harness />);
 
     await user.click(screen.getByTestId("configure-automatically-button"));
 
     expect(screen.getByTestId("auto-quality-selector")).toBeInTheDocument();
-    expect(screen.queryByTestId("auto-objective-selector")).not.toBeInTheDocument();
+    expect(screen.getByTestId("auto-objective-selector")).toBeInTheDocument();
     expect(screen.queryByTestId("template-selector")).not.toBeInTheDocument();
     expect(useAutoRouterRecommendation).toHaveBeenCalledWith(
-      expect.objectContaining({ qualityLevel: "balanced", enabled: true }),
+      expect.objectContaining({ qualityLevel: "balanced", optimizeFor: "balanced", enabled: true }),
     );
   });
 
@@ -195,12 +195,38 @@ describe("AddAutoRouterTab", () => {
     expect(screen.queryByTestId("auto-quality-selector")).not.toBeInTheDocument();
   });
 
-  it("submits the editable static config generated for the caller's available model groups", async () => {
+  it("submits the editable policy generated for the caller's available model groups", async () => {
     const user = userEvent.setup();
+    const autoSetup: AutoSetupConfigPayload = {
+      schema_version: "1" as const,
+      snapshot_id: "snapshot-test",
+      snapshot_sha256: "a".repeat(64),
+      quality_level: "max" as const,
+      optimize_for: "task_completion_speed" as const,
+      tier_policies: Object.fromEntries(
+        ["SIMPLE", "MEDIUM", "COMPLEX", "REASONING"].map((tier) => [
+          tier,
+          {
+            selection_mode: tier === "SIMPLE" || tier === "MEDIUM" ? "runtime_response_latency" : "snapshot_ranked",
+            candidates: [
+              {
+                model_name: "smart-group",
+                benchmark_model_id: "benchmark-smart",
+                quality_lower_bound: 0.98,
+                cost_per_completed_task_usd: 0.04,
+              },
+            ],
+            ...(tier === "SIMPLE" || tier === "MEDIUM" ? { cold_start_model: "smart-group" } : {}),
+            evidence_status: "test",
+          },
+        ]),
+      ),
+    };
     mockFetchAvailableModels.mockResolvedValue([{ model_group: "smart-group", mode: "chat" }]);
     const recommendationQuery = {
       data: {
         quality_level: "max",
+        optimize_for: "task_completion_speed",
         snapshot_id: "snapshot-test",
         snapshot_generated_at: "2026-09-03T00:00:00Z",
         available_model_group_count: 2,
@@ -219,6 +245,7 @@ describe("AddAutoRouterTab", () => {
           deployment_affinity: false,
           modality_routing: false,
           modality_pin_override: false,
+          auto_setup: autoSetup,
         },
       },
       isPending: false,
@@ -236,17 +263,9 @@ describe("AddAutoRouterTab", () => {
     await user.click(screen.getByRole("button", { name: /add auto router/i }));
 
     await waitFor(() => expect(handleAddAutoRouterSubmit).toHaveBeenCalled());
-    const submitted = vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0].complexity_router_config;
-    expect(submitted).toMatchObject({
-      tiers: {
-        SIMPLE: ["smart-group"],
-        MEDIUM: ["smart-group"],
-        COMPLEX: ["smart-group"],
-        REASONING: ["smart-group"],
-      },
-      classifier_type: "heuristic_v2",
+    expect(vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0].complexity_router_config).toMatchObject({
+      auto_setup: autoSetup,
     });
-    expect(submitted).not.toHaveProperty("auto_setup");
   });
 
   // Detailed Configuration starts collapsed after manual setup is selected; a caller opts into the
diff --git a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx
index cf3dc7d2a58..b5d3703b0f7 100644
--- a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx
+++ b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx
@@ -30,11 +30,12 @@ import ComplexityRouterConfig, {
   DEFAULT_DEPLOYMENT_AFFINITY,
   DEFAULT_TIER_DISTANCE_PENALTY,
 } from "./ComplexityRouterConfig";
-import { getRecommendationBlockedReason, type SetupMode } from "./auto_setup_policy";
+import { detachAutoSetupAfterPoolEdit, getRecommendationBlockedReason, type SetupMode } from "./auto_setup_policy";
 import { KeywordTierRule } from "./KeywordTierRules";
 import { DEFAULT_ESCALATION_KEYWORDS } from "./EscalationKeywords";
 import { DEFAULT_MATCH_THRESHOLD } from "./SemanticKeywordMatching";
 import {
+  AutoSetupObjective,
   AutoSetupQualityLevel,
   BuildComplexityRouterConfigParams,
   buildComplexityRouterConfig,
@@ -187,6 +188,7 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
   const [modelAccessGroups, setModelAccessGroups] = useState<string[]>([]);
   const [setupMode, setSetupMode] = useState<SetupMode>("manual");
   const [qualityLevel, setQualityLevel] = useState<AutoSetupQualityLevel>("balanced");
+  const [optimizeFor, setOptimizeFor] = useState<AutoSetupObjective>("balanced");
   const appliedRecommendationKey = useRef<string | null>(null);
 
   const [complexityRouterConfig, setComplexityRouterConfig] = useState<ComplexityRouterConfigValue>({
@@ -278,6 +280,7 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
   const recommendationParams = {
     accessToken,
     qualityLevel,
+    optimizeFor,
     teamId: requiresTeamScope ? watchedTeamId : undefined,
     enabled: recommendationEnabled,
   };
@@ -354,6 +357,7 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
     const key = [
       recommendation.snapshot_id,
       qualityLevel,
+      optimizeFor,
       watchedTeamId ?? "",
       ...recommendation.matched_model_groups,
     ].join(":");
@@ -362,7 +366,7 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
     setSelectedPreset(undefined);
     applyPrefill(buildPresetPrefill(recommendation.complexity_router_config, availability));
     setDetailsExpanded(false);
-  }, [recommendation, setupMode, qualityLevel, watchedTeamId, availability, applyPrefill]);
+  }, [recommendation, setupMode, qualityLevel, optimizeFor, watchedTeamId, availability, applyPrefill]);
 
   const handlePresetChange = (presetKey: string | undefined) => {
     if (!presetKey || presetKey === "custom") {
@@ -401,6 +405,11 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
     hasRecommendation: recommendation !== undefined,
   };
   const recommendationBlockedReason = getRecommendationBlockedReason(recommendationStatus);
+  const regenerateAutoSetup = () => {
+    if (!recommendation) return;
+    applyPrefill(buildPresetPrefill(recommendation.complexity_router_config, availability));
+    setDetailsExpanded(false);
+  };
   const submitBlockedReason =
     recommendationBlockedReason ??
     getSubmitBlockedReason(
@@ -449,6 +458,7 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
     reasoningOverrideMinScore: complexityRouterConfig.reasoning_override_min_score,
     enableContextWindowEscalation: complexityRouterConfig.enable_context_window_escalation,
     contextWindowEscalationBuffer: complexityRouterConfig.context_window_escalation_buffer,
+    autoSetup: complexityRouterConfig.auto_setup,
   };
 
   const submitRecommendedRouter = async (name: string) => {
@@ -588,14 +598,18 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
               {setupMode === "auto" ? (
                 <AutoSetupControls
                   qualityLevel={qualityLevel}
+                  optimizeFor={optimizeFor}
                   onQualityLevelChange={setQualityLevel}
+                  onOptimizeForChange={setOptimizeFor}
                   recommendation={recommendation}
                   recommendationEnabled={recommendationEnabled}
                   recommendationPending={recommendationPending}
                   recommendationFailed={recommendationFailed}
                   recommendationError={recommendationError}
                   requiresTeamScope={requiresTeamScope}
+                  hasAutoPolicy={Boolean(complexityRouterConfig.auto_setup)}
                   onRetry={() => void refetchRecommendation()}
+                  onRegenerate={regenerateAutoSetup}
                   onBack={returnToManualSetup}
                 />
               ) : (
@@ -700,7 +714,9 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
                       onEditingTiersChange={setEditingTiers}
                       modelInfo={modelInfo}
                       value={complexityRouterConfig}
-                      onChange={setComplexityRouterConfig}
+                      onChange={(next) =>
+                        setComplexityRouterConfig((previous) => detachAutoSetupAfterPoolEdit(previous, next))
+                      }
                       customTechnicalKeywords={customTechnicalKeywords}
                       onCustomTechnicalKeywordsChange={setCustomTechnicalKeywords}
                       keywordTierRules={keywordTierRules}
diff --git a/ui/litellm-dashboard/src/components/add_model/auto_setup_policy.ts b/ui/litellm-dashboard/src/components/add_model/auto_setup_policy.ts
index 08efd9ecad0..1ef3303d124 100644
--- a/ui/litellm-dashboard/src/components/add_model/auto_setup_policy.ts
+++ b/ui/litellm-dashboard/src/components/add_model/auto_setup_policy.ts
@@ -1,3 +1,5 @@
+import type { ComplexityRouterConfigValue } from "./ComplexityRouterConfig";
+
 export type SetupMode = "auto" | "manual";
 
 export interface RecommendationStatus {
@@ -23,3 +25,18 @@ export const getRecommendationBlockedReason = ({
   }
   return null;
 };
+
+const autoSetupPoolShape = (value: ComplexityRouterConfigValue): string =>
+  JSON.stringify({
+    tiers: value.tiers,
+    customTierSet: value.custom_tier_set,
+    tierModelParams: value.tier_model_params,
+  });
+
+export const detachAutoSetupAfterPoolEdit = (
+  previous: ComplexityRouterConfigValue,
+  next: ComplexityRouterConfigValue,
+): ComplexityRouterConfigValue =>
+  previous.auto_setup && autoSetupPoolShape(previous) !== autoSetupPoolShape(next)
+    ? { ...next, auto_setup: undefined }
+    : next;
diff --git a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts
index 2a8454ae246..446dd2f7b61 100644
--- a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts
+++ b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts
@@ -138,6 +138,7 @@ export interface BuildComplexityRouterConfigParams {
   tierModelParams?: TierModelParamsByTier;
   enableContextWindowEscalation?: boolean;
   contextWindowEscalationBuffer?: number;
+  autoSetup?: AutoSetupConfigPayload;
 }
 
 /**
@@ -196,9 +197,34 @@ export interface ComplexityRouterConfigPayload {
   enable_context_window_escalation?: boolean;
   context_window_escalation_buffer?: number;
   tier_model_configs?: Record<string, { model_name: string; litellm_params: TierModelParams }[]>;
+  auto_setup?: AutoSetupConfigPayload;
 }
 
 export type AutoSetupQualityLevel = "economy" | "balanced" | "high" | "max";
+export type AutoSetupObjective = "cost" | "task_completion_speed" | "balanced";
+
+export interface AutoSetupCandidatePayload {
+  model_name: string;
+  benchmark_model_id: string;
+  quality_lower_bound: number;
+  cost_per_completed_task_usd: number;
+}
+
+export interface AutoSetupTierPolicyPayload {
+  selection_mode: "snapshot_ranked" | "runtime_response_latency";
+  candidates: AutoSetupCandidatePayload[];
+  cold_start_model?: string | null;
+  evidence_status: string;
+}
+
+export interface AutoSetupConfigPayload {
+  schema_version: "1";
+  snapshot_id: string;
+  snapshot_sha256: string;
+  quality_level: AutoSetupQualityLevel;
+  optimize_for: AutoSetupObjective;
+  tier_policies: Record<string, AutoSetupTierPolicyPayload>;
+}
 
 export const serializeTierLabels = (tierLabels: ComplexityTierLabels | undefined): ComplexityTierLabels | undefined => {
   const renamed = TIER_KEYS.map((tier) => [tier, tierLabels?.[tier]?.trim() ?? ""] as const).filter(
@@ -458,6 +484,7 @@ export const buildComplexityRouterConfig = ({
   tierModelParams,
   enableContextWindowEscalation,
   contextWindowEscalationBuffer,
+  autoSetup,
 }: BuildComplexityRouterConfigParams): ComplexityRouterConfigPayload => {
   const serializedTierModelConfigs = customTierSet
     ? serializeTierModelConfigs(
@@ -524,6 +551,7 @@ export const buildComplexityRouterConfig = ({
     ...(contextWindowEscalationBuffer !== undefined && {
       context_window_escalation_buffer: contextWindowEscalationBuffer,
     }),
+    ...(autoSetup && { auto_setup: autoSetup }),
     ...scorerKnobs,
   };
   if (!customTierSet) return payload;
diff --git a/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts b/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts
index f24f3033901..8129d91de30 100644
--- a/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts
+++ b/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts
@@ -553,6 +553,7 @@ describe("managed keys survive an untouched open-and-save", () => {
     "fallback_tier",
     "classification_prompt",
     "hybrid_boundary_margin",
+    "auto_setup",
   ]);
 
   it("carries every managed key a built-in router can hold through hydrate then save", () => {
@@ -565,6 +566,21 @@ describe("managed keys survive an untouched open-and-save", () => {
     expect(dropped).toEqual([]);
   });
 
+  it("round-trips the stored Auto policy while its model pools stay unchanged", () => {
+    const autoSetup = {
+      schema_version: "1" as const,
+      snapshot_id: "snapshot-test",
+      snapshot_sha256: "a".repeat(64),
+      quality_level: "max" as const,
+      optimize_for: "cost" as const,
+      tier_policies: {},
+    };
+    const stored = { ...STORED, classifier_type: "heuristic" as const, auto_setup: autoSetup };
+    const hydrated = hydrateComplexityRouterConfig(stored, undefined);
+
+    expect(buildUpdatedComplexityRouterConfig(stored, hydrated).auto_setup).toEqual(autoSetup);
+  });
+
   it("drops a stored local-scorer threshold when the operator converts the router to custom tiers", () => {
     const hydrated = hydrateComplexityRouterConfig(STORED_ALL_MANAGED, undefined);
     const converted = {
diff --git a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx
index 54766df93ec..9d733d14d8b 100644
--- a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx
+++ b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx
@@ -38,6 +38,7 @@ import {
   hydratePlanModeMinTier,
   hydrateTierLabels,
   dryRunRejection,
+  type AutoSetupConfigPayload,
 } from "../add_model/build_complexity_router_config";
 import { KeywordTierRule } from "../add_model/KeywordTierRules";
 import { DEFAULT_MATCH_THRESHOLD } from "../add_model/SemanticKeywordMatching";
@@ -60,6 +61,7 @@ import ComplexityRouterConfig, {
   DEFAULT_DEPLOYMENT_AFFINITY,
   DEFAULT_TIER_DISTANCE_PENALTY,
 } from "../add_model/ComplexityRouterConfig";
+import { detachAutoSetupAfterPoolEdit } from "../add_model/auto_setup_policy";
 import {
   Dialog,
   DialogContent,
@@ -114,6 +116,7 @@ export interface StoredComplexityRouterConfig {
   return_raw_model_name?: boolean;
   enable_context_window_escalation?: unknown;
   context_window_escalation_buffer?: unknown;
+  auto_setup?: AutoSetupConfigPayload;
 }
 
 /**
@@ -202,6 +205,7 @@ export const hydrateComplexityRouterConfig = (
       typeof parsedConfig.context_window_escalation_buffer === "number"
         ? parsedConfig.context_window_escalation_buffer
         : undefined,
+    auto_setup: parsedConfig.auto_setup,
   };
 };
 
@@ -238,6 +242,7 @@ export const MANAGED_COMPLEXITY_ROUTER_KEYS = new Set([
   "reasoning_override_min_score",
   "enable_context_window_escalation",
   "context_window_escalation_buffer",
+  "auto_setup",
 ]);
 
 // Managed only when the caller passes the corresponding state. A caller that does not render
@@ -343,6 +348,7 @@ export const buildUpdatedComplexityRouterConfig = (
     tierModelParams: value.tier_model_params,
     enableContextWindowEscalation: value.enable_context_window_escalation,
     contextWindowEscalationBuffer: value.context_window_escalation_buffer,
+    autoSetup: value.auto_setup,
   };
   const built = buildComplexityRouterConfig(builderParams);
 
@@ -723,7 +729,7 @@ const EditAutoRouterModal: React.FC<EditAutoRouterModalProps> = ({
                     modelInfo={modelInfo}
                     value={complexityRouterConfig}
                     onChange={(config) => {
-                      setComplexityRouterConfig(config);
+                      setComplexityRouterConfig((previous) => detachAutoSetupAfterPoolEdit(previous, config));
                     }}
                     customTechnicalKeywords={customTechnicalKeywords}
                     onCustomTechnicalKeywordsChange={setCustomTechnicalKeywords}
diff --git a/ui/litellm-dashboard/src/components/networking.tsx b/ui/litellm-dashboard/src/components/networking.tsx
index cb0f88b1d43..00ee0986c43 100644
--- a/ui/litellm-dashboard/src/components/networking.tsx
+++ b/ui/litellm-dashboard/src/components/networking.tsx
@@ -89,7 +89,11 @@ import type {
   CoordinationRedisTestResponse,
 } from "@/app/(dashboard)/caching/_components/coordination_redis_settings/types";
 import { MCP_TOOLS_PREVIEW_FORBIDDEN_MESSAGE } from "./mcp_tools/constants";
-import type { AutoSetupQualityLevel, ComplexityRouterConfigPayload } from "./add_model/build_complexity_router_config";
+import type {
+  AutoSetupObjective,
+  AutoSetupQualityLevel,
+  ComplexityRouterConfigPayload,
+} from "./add_model/build_complexity_router_config";
 import type { AutoRouterPresetsResponse } from "@/lib/autorouter_presets";
 import type { VectorStoreIndex } from "@/app/(dashboard)/vector-stores/_components/IndexesTab";
 import type { RoutingDecision } from "./view_logs/LogDetailsDrawer/RoutingDecisionCard";
@@ -2418,13 +2422,14 @@ export interface AutoRouterRoutingTestRequest {
 
 export interface AutoRouterRecommendationResponse {
   quality_level: AutoSetupQualityLevel;
+  optimize_for: AutoSetupObjective;
   snapshot_id: string;
   snapshot_generated_at: string;
   available_model_group_count: number;
   matched_model_groups: string[];
   excluded_model_groups: Array<{
     model_group: string;
-    reason: "no_benchmark_match" | "mixed_model_group";
+    reason: "no_benchmark_match" | "mixed_model_group" | "pricing_unavailable";
   }>;
   complexity_router_config: ComplexityRouterConfigPayload;
 }
@@ -2432,12 +2437,14 @@ export interface AutoRouterRecommendationResponse {
 export const getAutoRouterRecommendation = async (
   accessToken: string,
   qualityLevel: AutoSetupQualityLevel,
+  optimizeFor: AutoSetupObjective,
   teamId?: string,
 ): Promise<AutoRouterRecommendationResponse> =>
   await apiClient.get<AutoRouterRecommendationResponse>("/auto_router/recommendation", {
     accessToken,
     query: {
       quality_level: qualityLevel,
+      optimize_for: optimizeFor,
       ...(teamId ? { team_id: teamId } : {}),
     },
   });
diff --git a/ui/litellm-dashboard/src/lib/autorouter_presets.ts b/ui/litellm-dashboard/src/lib/autorouter_presets.ts
index ff482bd23b8..4f80abb19c0 100644
--- a/ui/litellm-dashboard/src/lib/autorouter_presets.ts
+++ b/ui/litellm-dashboard/src/lib/autorouter_presets.ts
@@ -294,6 +294,7 @@ export const buildPresetPrefill = (
       return_raw_model_name: config.return_raw_model_name,
       enable_context_window_escalation: config.enable_context_window_escalation,
       context_window_escalation_buffer: config.context_window_escalation_buffer,
+      auto_setup: config.auto_setup,
     },
     customTechnicalKeywords: config.custom_technical_keywords ?? [],
     keywordTierRules: hydrateKeywordTierRules(config.keyword_tier_rules ?? []),
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 02dbfb98b5d..b9a181bde44 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -23449,7 +23449,7 @@ export interface components {
              * Reason
              * @enum {string}
              */
-            reason: "no_benchmark_match" | "mixed_model_group";
+            reason: "no_benchmark_match" | "mixed_model_group" | "pricing_unavailable";
         };
         /**
          * AutoRouterPresetConfig
@@ -23509,6 +23509,11 @@ export interface components {
             excluded_model_groups: components["schemas"]["AutoRouterExcludedModelGroup"][];
             /** Matched Model Groups */
             matched_model_groups: string[];
+            /**
+             * Optimize For
+             * @enum {string}
+             */
+            optimize_for: "cost" | "task_completion_speed" | "balanced";
             /**
              * Quality Level
              * @enum {string}
@@ -23592,6 +23597,67 @@ export interface components {
             /** @description The decision record this request would have written to its log row */
             routing_decision: components["schemas"]["StandardLoggingRoutingDecision"];
         };
+        /**
+         * AutoSetupCandidate
+         * @description One benchmark-backed model group in an Auto-generated tier policy.
+         */
+        AutoSetupCandidate: {
+            /** Benchmark Model Id */
+            benchmark_model_id: string;
+            /** Cost Per Completed Task Usd */
+            cost_per_completed_task_usd: number;
+            /** Model Name */
+            model_name: string;
+            /** Quality Lower Bound */
+            quality_lower_bound: number;
+        };
+        /**
+         * AutoSetupConfig
+         * @description Provenance and selection policy stamped only by zero-setup Auto creation.
+         */
+        AutoSetupConfig: {
+            /**
+             * Optimize For
+             * @enum {string}
+             */
+            optimize_for: "cost" | "task_completion_speed" | "balanced";
+            /**
+             * Quality Level
+             * @enum {string}
+             */
+            quality_level: "economy" | "balanced" | "high" | "max";
+            /**
+             * Schema Version
+             * @default 1
+             * @constant
+             */
+            schema_version: "1";
+            /** Snapshot Id */
+            snapshot_id: string;
+            /** Snapshot Sha256 */
+            snapshot_sha256: string;
+            /** Tier Policies */
+            tier_policies: {
+                [key: string]: components["schemas"]["AutoSetupTierPolicy"];
+            };
+        };
+        /**
+         * AutoSetupTierPolicy
+         * @description How an Auto-generated tier chooses from its quality-admitted model groups.
+         */
+        AutoSetupTierPolicy: {
+            /** Candidates */
+            candidates: components["schemas"]["AutoSetupCandidate"][];
+            /** Cold Start Model */
+            cold_start_model?: string | null;
+            /** Evidence Status */
+            evidence_status: string;
+            /**
+             * Selection Mode
+             * @enum {string}
+             */
+            selection_mode: "snapshot_ranked" | "runtime_response_latency";
+        };
         /** BaseLitellmParams */
         BaseLitellmParams: {
             /**
@@ -34656,6 +34722,8 @@ export interface components {
             adaptive_eligible: "all" | "classified_tier";
             /** @description Quality vs cost weights for adaptive selection (used when adaptive=True) */
             adaptive_weights?: components["schemas"]["AdaptiveRouterWeights"];
+            /** @description Benchmark snapshot and per-tier selection policy generated by the zero-setup Auto flow. Absent on manually configured routers, whose pool selection behavior is unchanged */
+            auto_setup?: components["schemas"]["AutoSetupConfig"] | null;
             /**
              * Classification Mode
              * @description When to run the complexity classifier. 'every_request' (the default) classifies every inference request, including the tool-result continuation turns of an agentic loop. 'user_turn' classifies only requests whose newest turn is a new human ask and replays the session's held routing decision on continuation turns, which cuts classifier spend and eliminates mid-loop model switches. Continuations with no held decision to replay (no resolvable session_id, expired pin, fresh restart) still classify. Unlike session_affinity, a new human ask always re-classifies, so a session can still move tiers between asks. Suppressed when plugins are configured, for the same reason session_affinity is: a replayed decision would bypass the plugin pipeline.
@@ -36012,6 +36080,23 @@ export interface components {
          * @description Per-request provenance for a pre-routing strategy (auto-router) decision.
          */
         StandardLoggingRoutingDecision: {
+            /**
+             * Auto Setup Optimize For
+             * @enum {string}
+             */
+            auto_setup_optimize_for?: "cost" | "task_completion_speed" | "balanced";
+            /**
+             * Auto Setup Quality Level
+             * @enum {string}
+             */
+            auto_setup_quality_level?: "economy" | "balanced" | "high" | "max";
+            /**
+             * Auto Setup Selection Mode
+             * @enum {string}
+             */
+            auto_setup_selection_mode?: "snapshot_ranked" | "runtime_response_latency";
+            /** Auto Setup Snapshot Id */
+            auto_setup_snapshot_id?: string;
             /**
              * Cause
              * @enum {string}
@@ -41215,6 +41300,8 @@ export interface operations {
             query?: {
                 /** @description How close every admitted model must be to the best available quality score */
                 quality_level?: "economy" | "balanced" | "high" | "max";
+                /** @description What to minimize after the selected quality gate is applied */
+                optimize_for?: "cost" | "task_completion_speed" | "balanced";
                 /** @description Team whose model access the recommendation must use */
                 team_id?: string | null;
             };