Merge 01fafb270c into 461a58c40a

litellm-proxy-extras/litellm_proxy_extras/migrations/20261003140000_add_timed_completion_tokens/migration.sql

@@ -0,0 +1,6 @@
+ALTER TABLE "LiteLLM_DailyUserSpend" ADD COLUMN IF NOT EXISTS "timed_completion_tokens" BIGINT;
+ALTER TABLE "LiteLLM_DailyOrganizationSpend" ADD COLUMN IF NOT EXISTS "timed_completion_tokens" BIGINT;
+ALTER TABLE "LiteLLM_DailyEndUserSpend" ADD COLUMN IF NOT EXISTS "timed_completion_tokens" BIGINT;
+ALTER TABLE "LiteLLM_DailyAgentSpend" ADD COLUMN IF NOT EXISTS "timed_completion_tokens" BIGINT;
+ALTER TABLE "LiteLLM_DailyTeamSpend" ADD COLUMN IF NOT EXISTS "timed_completion_tokens" BIGINT;
+ALTER TABLE "LiteLLM_DailyTagSpend" ADD COLUMN IF NOT EXISTS "timed_completion_tokens" BIGINT;

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -870,6 +870,7 @@ model LiteLLM_DailyUserSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 
@@ -939,6 +940,7 @@ model LiteLLM_DailyOrganizationSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 
@@ -977,6 +979,7 @@ model LiteLLM_DailyEndUserSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
   @@unique([end_user_id, date, api_key, model, custom_llm_provider, mcp_namespaced_tool_name, endpoint])
@@ -1014,6 +1017,7 @@ model LiteLLM_DailyAgentSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
   @@unique([agent_id, date, api_key, model, custom_llm_provider, mcp_namespaced_tool_name, endpoint])
@@ -1051,6 +1055,7 @@ model LiteLLM_DailyTeamSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   ptu_flat_cost       Float    @default(0.0)
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
@@ -1091,6 +1096,7 @@ model LiteLLM_DailyTagSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 

litellm/integrations/otel/model/spans.py

@@ -348,6 +348,7 @@ _PRISMA_MODELS: Final[frozenset[str]] = frozenset(
         "LiteLLM_WorkflowRun",
         "LiteLLM_WorkflowEvent",
         "LiteLLM_WorkflowMessage",
+        "LiteLLM_BackgroundInteractionSettlement",
         "LiteLLM_Lens",
         "LiteLLM_LensRun",
         "LiteLLM_LensWorker",

litellm/proxy/_lazy_openapi_snapshot.json

@@ -4016,6 +4016,13 @@
             },
             "metrics": {
               "$ref": "#/components/schemas/SpendMetrics"
+            },
+            "provider_breakdown": {
+              "additionalProperties": {
+                "$ref": "#/components/schemas/ProviderThroughputMetrics"
+              },
+              "title": "Provider Breakdown",
+              "type": "object"
             }
           },
           "required": [
@@ -4343,6 +4350,44 @@
           "title": "PatchAgentRequest",
           "type": "object"
         },
+        "ProviderThroughputMetrics": {
+          "properties": {
+            "output_tokens_per_second": {
+              "anyOf": [
+                {
+                  "type": "number"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Output Tokens Per Second"
+            },
+            "timed_completion_tokens": {
+              "anyOf": [
+                {
+                  "type": "integer"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Timed Completion Tokens"
+            },
+            "timed_requests": {
+              "default": 0,
+              "title": "Timed Requests",
+              "type": "integer"
+            },
+            "total_response_time_ms": {
+              "default": 0,
+              "title": "Total Response Time Ms",
+              "type": "integer"
+            }
+          },
+          "title": "ProviderThroughputMetrics",
+          "type": "object"
+        },
         "SpendAnalyticsPaginatedResponse": {
           "properties": {
             "metadata": {

litellm/proxy/_types.py

@@ -5705,6 +5705,7 @@ class BaseDailySpendTransaction(TypedDict):
     failed_requests: int
     total_response_time_ms: NotRequired[int]  # writable-ok: the rollup queue accumulates into this key in place
     timed_requests: NotRequired[int]  # writable-ok: the rollup queue accumulates into this key in place
+    timed_completion_tokens: NotRequired[int | None]
 
 
 class DailyTeamSpendTransaction(BaseDailySpendTransaction):

litellm/proxy/db/daily_spend_bulk_upsert.py

@@ -91,6 +91,7 @@ _COUNTER_COLUMNS: Final = (
     "compression_saved_tokens",
     "total_response_time_ms",
     "timed_requests",
+    "timed_completion_tokens",
 )
 _SPEND_COLUMNS: Final = (
     "spend",
@@ -124,6 +125,23 @@ def _as_float(value: object) -> float:
     return float(value) if isinstance(value, (int, float)) else 0.0
 
 
+def _counter_total(column: str, group: Sequence[SpendRow]) -> int | None:
+    values: Final = tuple(row.get(column) for row in group)
+    has_unknown_timed_tokens: Final = column == "timed_completion_tokens" and any(
+        row.get("timed_completion_tokens") is None and _as_int(row.get("timed_requests")) > 0 for row in group
+    )
+    if has_unknown_timed_tokens:
+        return None
+    return sum(_as_int(value) for value in values)
+
+
+def _counter_value(column: str, transaction: SpendRow) -> int | None:
+    value: Final = transaction.get(column)
+    if column == "timed_completion_tokens" and value is None and _as_int(transaction.get("timed_requests")) > 0:
+        return None
+    return _as_int(value)
+
+
 def conflict_key(table: DailySpendTable, transaction: SpendRow) -> tuple[str, ...]:
     """The tuple the database arbitrates the upsert on, normalized free of NULLs."""
     return tuple(_as_text(transaction.get(column)) for column in (table.entity_id_column, *_KEY_COLUMNS))
@@ -134,7 +152,7 @@ def _merge(group: Sequence[SpendRow]) -> SpendRow:
         return group[0]
     return {
         **group[0],
-        **{column: sum(_as_int(row.get(column)) for row in group) for column in _COUNTER_COLUMNS},
+        **{column: _counter_total(column, group) for column in _COUNTER_COLUMNS},
         **{column: sum(_as_float(row.get(column)) for row in group) for column in _SPEND_COLUMNS},
     }
 
@@ -165,7 +183,7 @@ def _row_params(
         str(uuid.uuid4()),
         *key,
         None if transaction.get("model_group") is None else _as_text(transaction.get("model_group")),
-        *(_as_int(transaction.get(column)) for column in _COUNTER_COLUMNS),
+        *(_counter_value(column, transaction) for column in _COUNTER_COLUMNS),
         *(_as_float(transaction.get(column)) for column in _SPEND_COLUMNS),
         *((None if request_id is None else _as_text(request_id),) if table.carries_request_id else ()),
     )
@@ -199,9 +217,18 @@ def build_bulk_upsert(
         + ", (NOW() AT TIME ZONE 'UTC'))"
         for row_index in range(len(batch))
     )
-    increments: Final = ", ".join(
-        f'"{column}" = {quoted_table}."{column}" + EXCLUDED."{column}"'
-        for column in (*_COUNTER_COLUMNS, *_SPEND_COLUMNS)
+    regular_increment_columns: Final = tuple(
+        column for column in (*_COUNTER_COLUMNS, *_SPEND_COLUMNS) if column != "timed_completion_tokens"
+    )
+    regular_increments: Final = ", ".join(
+        f'"{column}" = {quoted_table}."{column}" + EXCLUDED."{column}"' for column in regular_increment_columns
+    )
+    timed_completion_tokens_increment: Final = (
+        f'"timed_completion_tokens" = CASE '
+        f'WHEN ({quoted_table}."timed_requests" > 0 AND {quoted_table}."timed_completion_tokens" IS NULL) '
+        'OR (EXCLUDED."timed_requests" > 0 AND EXCLUDED."timed_completion_tokens" IS NULL) THEN NULL '
+        f'ELSE COALESCE({quoted_table}."timed_completion_tokens", 0) '
+        '+ COALESCE(EXCLUDED."timed_completion_tokens", 0) END'
     )
     # request_id names one arbitrary contributing request, so an entry carrying none must
     # not blank out the one already recorded.
@@ -214,7 +241,7 @@ def build_bulk_upsert(
         f'INSERT INTO {quoted_table} ({_quoted(columns)}, "updated_at")\n'
         f"VALUES {rows}\n"
         f"ON CONFLICT ({_quoted((table.entity_id_column, *_KEY_COLUMNS))}) DO UPDATE SET\n"
-        f"  {increments}{request_id_update},\n"
+        f"  {regular_increments}, {timed_completion_tokens_increment}{request_id_update},\n"
         f"  \"updated_at\" = (NOW() AT TIME ZONE 'UTC')"
     )
     return sql, tuple(value for key, transaction in batch for value in _row_params(table, key, transaction))

litellm/proxy/db/db_spend_update_writer.py

@@ -2820,6 +2820,7 @@ class DBSpendUpdateWriter:
                 autorouter_savings_spend=0.0 if is_internal_call else savings_spend.autorouter,
                 total_response_time_ms=timed_duration_ms or 0,
                 timed_requests=0 if timed_duration_ms is None else 1,
+                timed_completion_tokens=0 if timed_duration_ms is None else cast(int, payload["completion_tokens"]),
             )
             return daily_transaction
         except Exception as e:

litellm/proxy/db/db_transaction_queue/daily_spend_update_queue.py

@@ -1,6 +1,8 @@
 import asyncio
-from collections.abc import Coroutine
+from collections.abc import Coroutine, Iterator
 from copy import deepcopy
+from functools import reduce
+from itertools import groupby
 from typing import Final
 
 from litellm._logging import verbose_proxy_logger
@@ -13,6 +15,62 @@ from litellm.proxy.db.db_transaction_queue.base_update_queue import (
 from litellm.types.services import ServiceTypes
 
 
+def _daily_spend_updates(
+    updates: list[dict[str, BaseDailySpendTransaction]],
+) -> Iterator[tuple[str, BaseDailySpendTransaction]]:
+    for update in updates:
+        yield from update.items()
+
+
+def _sum_timed_completion_tokens(
+    left_tokens: int | None,
+    left_requests: int,
+    right_tokens: int | None,
+    right_requests: int,
+) -> int | None:
+    if (left_tokens is None and left_requests > 0) or (right_tokens is None and right_requests > 0):
+        return None
+    return (left_tokens or 0) + (right_tokens or 0)
+
+
+def _merge_daily_spend_transactions(
+    existing: BaseDailySpendTransaction,
+    payload: BaseDailySpendTransaction,
+) -> BaseDailySpendTransaction:
+    return {
+        **existing,
+        "spend": existing["spend"] + payload["spend"],
+        "prompt_tokens": existing["prompt_tokens"] + payload["prompt_tokens"],
+        "completion_tokens": existing["completion_tokens"] + payload["completion_tokens"],
+        "api_requests": existing["api_requests"] + payload["api_requests"],
+        "successful_requests": existing["successful_requests"] + payload["successful_requests"],
+        "failed_requests": existing["failed_requests"] + payload["failed_requests"],
+        "cache_read_input_tokens": (existing.get("cache_read_input_tokens", 0) or 0)
+        + (payload.get("cache_read_input_tokens", 0) or 0),
+        "cache_creation_input_tokens": (existing.get("cache_creation_input_tokens", 0) or 0)
+        + (payload.get("cache_creation_input_tokens", 0) or 0),
+        "compression_saved_tokens": (existing.get("compression_saved_tokens", 0) or 0)
+        + (payload.get("compression_saved_tokens", 0) or 0),
+        "compression_savings_spend": (existing.get("compression_savings_spend", 0) or 0)
+        + (payload.get("compression_savings_spend", 0) or 0),
+        "prompt_caching_savings_spend": (existing.get("prompt_caching_savings_spend", 0) or 0)
+        + (payload.get("prompt_caching_savings_spend", 0) or 0),
+        "gateway_injected_caching_savings_spend": (existing.get("gateway_injected_caching_savings_spend", 0) or 0)
+        + (payload.get("gateway_injected_caching_savings_spend", 0) or 0),
+        "autorouter_savings_spend": (existing.get("autorouter_savings_spend", 0) or 0)
+        + (payload.get("autorouter_savings_spend", 0) or 0),
+        "total_response_time_ms": (existing.get("total_response_time_ms", 0) or 0)
+        + (payload.get("total_response_time_ms", 0) or 0),
+        "timed_requests": (existing.get("timed_requests", 0) or 0) + (payload.get("timed_requests", 0) or 0),
+        "timed_completion_tokens": _sum_timed_completion_tokens(
+            existing.get("timed_completion_tokens"),
+            existing.get("timed_requests", 0) or 0,
+            payload.get("timed_completion_tokens"),
+            payload.get("timed_requests", 0) or 0,
+        ),
+    }
+
+
 class DailySpendUpdateQueue(BaseUpdateQueue):
     """
     In memory buffer for daily spend updates that should be committed to the database
@@ -113,58 +171,14 @@ class DailySpendUpdateQueue(BaseUpdateQueue):
         updates: list[dict[str, BaseDailySpendTransaction]],
     ) -> dict[str, BaseDailySpendTransaction]:
         """Aggregate updates by daily_transaction_key."""
-        aggregated_daily_spend_update_transactions: Final[dict[str, BaseDailySpendTransaction]] = {}
-        for _update in updates:
-            for _key, payload in _update.items():
-                if _key in aggregated_daily_spend_update_transactions:
-                    daily_transaction = aggregated_daily_spend_update_transactions[_key]
-                    daily_transaction["spend"] += payload["spend"]
-                    daily_transaction["prompt_tokens"] += payload["prompt_tokens"]
-                    daily_transaction["completion_tokens"] += payload["completion_tokens"]
-                    daily_transaction["api_requests"] += payload["api_requests"]
-                    daily_transaction["successful_requests"] += payload["successful_requests"]
-                    daily_transaction["failed_requests"] += payload["failed_requests"]
-
-                    # Add optional metrics cache_read_input_tokens and cache_creation_input_tokens
-                    daily_transaction["cache_read_input_tokens"] = (
-                        payload.get("cache_read_input_tokens", 0) or 0
-                    ) + daily_transaction.get("cache_read_input_tokens", 0)
-
-                    daily_transaction["cache_creation_input_tokens"] = (
-                        payload.get("cache_creation_input_tokens", 0) or 0
-                    ) + daily_transaction.get("cache_creation_input_tokens", 0)
-
-                    daily_transaction["compression_saved_tokens"] = (
-                        payload.get("compression_saved_tokens", 0) or 0
-                    ) + daily_transaction.get("compression_saved_tokens", 0)
-
-                    daily_transaction["compression_savings_spend"] = (
-                        payload.get("compression_savings_spend", 0) or 0
-                    ) + daily_transaction.get("compression_savings_spend", 0)
-
-                    daily_transaction["prompt_caching_savings_spend"] = (
-                        payload.get("prompt_caching_savings_spend", 0) or 0
-                    ) + daily_transaction.get("prompt_caching_savings_spend", 0)
-
-                    daily_transaction["gateway_injected_caching_savings_spend"] = (
-                        payload.get("gateway_injected_caching_savings_spend", 0) or 0
-                    ) + daily_transaction.get("gateway_injected_caching_savings_spend", 0)
-
-                    daily_transaction["autorouter_savings_spend"] = (
-                        payload.get("autorouter_savings_spend", 0) or 0
-                    ) + daily_transaction.get("autorouter_savings_spend", 0)
-
-                    daily_transaction["total_response_time_ms"] = (
-                        payload.get("total_response_time_ms", 0) or 0
-                    ) + daily_transaction.get("total_response_time_ms", 0)
-
-                    daily_transaction["timed_requests"] = (
-                        payload.get("timed_requests", 0) or 0
-                    ) + daily_transaction.get("timed_requests", 0)
-
-                else:
-                    aggregated_daily_spend_update_transactions[_key] = deepcopy(payload)
-        return aggregated_daily_spend_update_transactions
+        ordered_updates: Final = sorted(_daily_spend_updates(updates), key=lambda update: update[0])
+        return {
+            key: reduce(
+                _merge_daily_spend_transactions,
+                (deepcopy(payload) for _, payload in grouped_updates),
+            )
+            for key, grouped_updates in groupby(ordered_updates, key=lambda update: update[0])
+        }
 
     async def _emit_new_item_added_to_queue_event(
         self,

litellm/proxy/management_endpoints/common_daily_activity.py

@@ -31,6 +31,7 @@ from litellm.types.proxy.management_endpoints.common_daily_activity import (
     KeyMetadata,
     KeyMetricWithMetadata,
     MetricWithMetadata,
+    ProviderThroughputMetrics,
     SpendAnalyticsPaginatedResponse,
     SpendMetrics,
 )
@@ -160,6 +161,9 @@ class DailySpendRecord(Protocol):
     @property
     def timed_requests(self) -> int: ...
 
+    @property
+    def timed_completion_tokens(self) -> int | None: ...
+
 
 class _KeyMetadataDict(TypedDict, total=False):
     key_alias: ReadOnly[str | None]
@@ -236,6 +240,56 @@ def update_metrics(existing_metrics: SpendMetrics, record: DailySpendRecord) ->
     return existing_metrics
 
 
+def _provider_throughput(
+    timed_completion_tokens: int | None,
+    total_response_time_ms: int,
+    timed_requests: int,
+) -> ProviderThroughputMetrics:
+    output_tokens_per_second: Final = (
+        timed_completion_tokens * 1000 / total_response_time_ms
+        if timed_completion_tokens is not None and timed_requests > 0 and total_response_time_ms > 0
+        else None
+    )
+    return ProviderThroughputMetrics(
+        timed_completion_tokens=timed_completion_tokens,
+        total_response_time_ms=total_response_time_ms,
+        timed_requests=timed_requests,
+        output_tokens_per_second=output_tokens_per_second,
+    )
+
+
+def _combined_timed_completion_tokens(
+    existing: ProviderThroughputMetrics | None,
+    current: int | None,
+) -> int | None:
+    if existing is None:
+        return current
+    if existing.timed_completion_tokens is None or current is None:
+        return None
+    return existing.timed_completion_tokens + current
+
+
+def _update_provider_throughput(
+    target: MetricWithMetadata,
+    provider: str,
+    record: DailySpendRecord,
+) -> None:
+    existing: Final = target.provider_breakdown.get(provider)
+    timed_completion_tokens: Final = _combined_timed_completion_tokens(
+        existing,
+        record.timed_completion_tokens,
+    )
+    target.provider_breakdown = {
+        **target.provider_breakdown,
+        provider: _provider_throughput(
+            timed_completion_tokens=timed_completion_tokens,
+            total_response_time_ms=(existing.total_response_time_ms if existing is not None else 0)
+            + (record.total_response_time_ms or 0),
+            timed_requests=(existing.timed_requests if existing is not None else 0) + (record.timed_requests or 0),
+        ),
+    }
+
+
 def _is_user_agent_tag(tag: str | None) -> bool:
     """Determine whether a tag should be treated as a User-Agent tag."""
     if not tag:
@@ -312,6 +366,12 @@ def update_breakdown_metrics(
         breakdown.models[model_key].metrics = update_metrics(breakdown.models[model_key].metrics, record)
 
         if not is_ptu_sentinel:
+            _update_provider_throughput(
+                breakdown.models[model_key],
+                record.custom_llm_provider or "unknown",
+                record,
+            )
+
             # Update API key breakdown for this model
             if record.api_key not in breakdown.models[model_key].api_key_breakdown:
                 breakdown.models[model_key].api_key_breakdown[record.api_key] = KeyMetricWithMetadata(
@@ -336,6 +396,12 @@ def update_breakdown_metrics(
         )
 
         if not is_ptu_sentinel:
+            _update_provider_throughput(
+                breakdown.model_groups[model_group_key],
+                record.custom_llm_provider or "unknown",
+                record,
+            )
+
             # Update API key breakdown for this model
             if record.api_key not in breakdown.model_groups[model_group_key].api_key_breakdown:
                 breakdown.model_groups[model_group_key].api_key_breakdown[record.api_key] = KeyMetricWithMetadata(
@@ -697,8 +763,10 @@ _API_KEY_ROLLED_UP_BIT: Final = 32  # 0b0100000
 _GROUP_DATE_API_KEY: Final = 31  # 0b0011111
 _GROUP_DATE_MODEL: Final = 47  # 0b0101111
 _GROUP_DATE_MODEL_API_KEY: Final = 15  # 0b0001111
+_GROUP_DATE_MODEL_PROVIDER: Final = 43  # 0b0101011
 _GROUP_DATE_MODEL_GROUP: Final = 55  # 0b0110111
 _GROUP_DATE_MODEL_GROUP_API_KEY: Final = 23  # 0b0010111
+_GROUP_DATE_MODEL_GROUP_PROVIDER: Final = 51  # 0b0110011
 _GROUP_DATE_PROVIDER: Final = 59  # 0b0111011
 _GROUP_DATE_PROVIDER_API_KEY: Final = 27  # 0b0011011
 _GROUP_DATE_MCP: Final = 61  # 0b0111101
@@ -779,6 +847,35 @@ def _aggregate_grouping_sets_records_sync(
             metrics=metrics, metadata=_key_metadata(api_key_metadata, api_key)
         )
 
+    def assign_provider_breakdown(
+        target: dict[str, MetricWithMetadata],
+        parent_key: str,
+        provider: str,
+        record: GroupingSetsRow,
+    ) -> None:
+        parent: Final = target.get(parent_key)
+        if parent is None:
+            target[parent_key] = MetricWithMetadata(
+                metrics=SpendMetrics(),
+                metadata={},
+                provider_breakdown={
+                    provider: _provider_throughput(
+                        record.timed_completion_tokens,
+                        record.total_response_time_ms or 0,
+                        record.timed_requests or 0,
+                    )
+                },
+            )
+            return
+        parent.provider_breakdown = {
+            **parent.provider_breakdown,
+            provider: _provider_throughput(
+                record.timed_completion_tokens,
+                record.total_response_time_ms or 0,
+                record.timed_requests or 0,
+            ),
+        }
+
     for record in records:
         level = record.group_level
         metrics = _record_to_spend_metrics(record)
@@ -806,6 +903,14 @@ def _aggregate_grouping_sets_records_sync(
         elif level == _GROUP_DATE_MODEL_API_KEY:
             if record.model and record.api_key and not is_ptu_sentinel:
                 assign_api_key_breakdown(breakdown.models, record.model, record.api_key, metrics)
+        elif level == _GROUP_DATE_MODEL_PROVIDER:
+            if record.model:
+                assign_provider_breakdown(
+                    breakdown.models,
+                    record.model,
+                    record.custom_llm_provider or "unknown",
+                    record,
+                )
         elif level == _GROUP_DATE_MODEL_GROUP:
             if record.model_group:
                 assign_metric_with_metadata(breakdown.model_groups, record.model_group, metrics)
@@ -817,6 +922,14 @@ def _aggregate_grouping_sets_records_sync(
                     record.api_key,
                     metrics,
                 )
+        elif level == _GROUP_DATE_MODEL_GROUP_PROVIDER:
+            if record.model_group:
+                assign_provider_breakdown(
+                    breakdown.model_groups,
+                    record.model_group,
+                    record.custom_llm_provider or "unknown",
+                    record,
+                )
         elif level == _GROUP_DATE_PROVIDER:
             # Only PTU sentinel rows carry ptu_flat_cost and they have no provider, so at
             # this level the sentinel's cost would land under "unknown". Withholding the

litellm/proxy/schema.prisma

@@ -870,6 +870,7 @@ model LiteLLM_DailyUserSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 
@@ -939,6 +940,7 @@ model LiteLLM_DailyOrganizationSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 
@@ -977,6 +979,7 @@ model LiteLLM_DailyEndUserSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
   @@unique([end_user_id, date, api_key, model, custom_llm_provider, mcp_namespaced_tool_name, endpoint])
@@ -1014,6 +1017,7 @@ model LiteLLM_DailyAgentSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
   @@unique([agent_id, date, api_key, model, custom_llm_provider, mcp_namespaced_tool_name, endpoint])
@@ -1051,6 +1055,7 @@ model LiteLLM_DailyTeamSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   ptu_flat_cost       Float    @default(0.0)
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
@@ -1091,6 +1096,7 @@ model LiteLLM_DailyTagSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 

litellm/repositories/daily_activity_sql.py

@@ -129,7 +129,13 @@ def _rollup_metric_select(table: DailyActivityTable) -> str:
             SUM(successful_requests)::bigint AS successful_requests,
             SUM(failed_requests)::bigint AS failed_requests,
             SUM(total_response_time_ms)::bigint AS total_response_time_ms,
-            SUM(timed_requests)::bigint AS timed_requests"""
+            SUM(timed_requests)::bigint AS timed_requests,
+            CASE
+                WHEN COUNT(*) FILTER (
+                    WHERE timed_requests > 0 AND timed_completion_tokens IS NULL
+                ) > 0 THEN NULL::bigint
+                ELSE COALESCE(SUM(timed_completion_tokens), 0)::bigint
+            END AS timed_completion_tokens"""
 
 
 def _validate_api_key_limit(api_key_limit: int) -> None:
@@ -177,7 +183,9 @@ def build_aggregated_sql(scope: DailyActivityScope, *, api_key_limit: int) -> Sq
         GROUP BY GROUPING SETS (
             (date),
             (date, model),
+            (date, model, custom_llm_provider),
             (date, {_MODEL_GROUP_EXPR}),
+            (date, {_MODEL_GROUP_EXPR}, custom_llm_provider),
             (date, custom_llm_provider),
             (date, mcp_namespaced_tool_name),
             (date, endpoint),

litellm/types/proxy/management_endpoints/common_daily_activity.py

@@ -56,10 +56,18 @@ class KeyMetricWithMetadata(MetricBase):
     metadata: KeyMetadata = Field(default_factory=KeyMetadata)
 
 
+class ProviderThroughputMetrics(BaseModel):
+    timed_completion_tokens: int | None = Field(default=None)
+    total_response_time_ms: int = Field(default=0)
+    timed_requests: int = Field(default=0)
+    output_tokens_per_second: float | None = Field(default=None)
+
+
 class MetricWithMetadata(MetricBase):
     metadata: dict[str, Any] = Field(default_factory=dict)
     # API key breakdown for this metric (e.g., which API keys are using this MCP server)
     api_key_breakdown: dict[str, KeyMetricWithMetadata] = Field(default_factory=dict)  # api_key -> {metrics, metadata}
+    provider_breakdown: dict[str, ProviderThroughputMetrics] = Field(default_factory=dict)
 
 
 class BreakdownMetrics(BaseModel):

litellm/types/repositories/daily_activity.py

@@ -124,6 +124,7 @@ class RollupMetricsRow:
     failed_requests: int | None
     total_response_time_ms: int | None
     timed_requests: int | None
+    timed_completion_tokens: int | None
 
 
 @dataclass(frozen=True, slots=True)
@@ -184,6 +185,7 @@ class DailyActivityRow(Protocol):
     failed_requests: int
     total_response_time_ms: int
     timed_requests: int
+    timed_completion_tokens: int | None
 
 
 @dataclass(frozen=True, slots=True)

schema.prisma

@@ -870,6 +870,7 @@ model LiteLLM_DailyUserSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 
@@ -939,6 +940,7 @@ model LiteLLM_DailyOrganizationSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 
@@ -977,6 +979,7 @@ model LiteLLM_DailyEndUserSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
   @@unique([end_user_id, date, api_key, model, custom_llm_provider, mcp_namespaced_tool_name, endpoint])
@@ -1014,6 +1017,7 @@ model LiteLLM_DailyAgentSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
   @@unique([agent_id, date, api_key, model, custom_llm_provider, mcp_namespaced_tool_name, endpoint])
@@ -1051,6 +1055,7 @@ model LiteLLM_DailyTeamSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   ptu_flat_cost       Float    @default(0.0)
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
@@ -1091,6 +1096,7 @@ model LiteLLM_DailyTagSpend {
   failed_requests     BigInt      @default(0)
   total_response_time_ms BigInt   @default(0)
   timed_requests      BigInt      @default(0)
+  timed_completion_tokens BigInt?
   created_at          DateTime @default(now())
   updated_at          DateTime @updatedAt
 

tests/unit/proxy/db/db_transaction_queue/test_daily_spend_update_queue.py

@@ -40,6 +40,7 @@ async def test_add_single_update(daily_spend_update_queue):
         "spend": 10.0,
         "prompt_tokens": 100,
         "completion_tokens": 50,
+        "timed_completion_tokens": 50,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -73,6 +74,7 @@ async def test_add_multiple_updates(daily_spend_update_queue):
         "spend": 5.0,
         "prompt_tokens": 200,
         "completion_tokens": 30,
+        "timed_completion_tokens": 0,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -105,9 +107,7 @@ async def test_add_multiple_updates(daily_spend_update_queue):
 @pytest.mark.asyncio
 async def test_aggregated_daily_spend_update_empty(daily_spend_update_queue):
     """Test aggregating updates from an empty queue"""
-    result = (
-        await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
-    )
+    result = await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
     assert result == {}
 
 
@@ -127,9 +127,7 @@ async def test_get_aggregated_daily_spend_update_transactions_single_key():
     updates = [{test_key: test_transaction}]
 
     # Test aggregation
-    result = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(
-        updates
-    )
+    result = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(updates)
 
     assert len(result) == 1
     assert test_key in result
@@ -162,9 +160,7 @@ async def test_get_aggregated_daily_spend_update_transactions_multiple_keys():
     updates = [{test_key1: test_transaction1}, {test_key2: test_transaction2}]
 
     # Test aggregation
-    result = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(
-        updates
-    )
+    result = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(updates)
 
     assert len(result) == 2
     assert test_key1 in result
@@ -181,6 +177,7 @@ async def test_get_aggregated_daily_spend_update_transactions_same_key():
         "spend": 10.0,
         "prompt_tokens": 100,
         "completion_tokens": 50,
+        "timed_completion_tokens": 50,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -190,6 +187,7 @@ async def test_get_aggregated_daily_spend_update_transactions_same_key():
         "spend": 5.0,
         "prompt_tokens": 200,
         "completion_tokens": 30,
+        "timed_completion_tokens": 0,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -199,6 +197,7 @@ async def test_get_aggregated_daily_spend_update_transactions_same_key():
         "spend": 15.0,  # 10 + 5
         "prompt_tokens": 300,  # 100 + 200
         "completion_tokens": 80,  # 50 + 30
+        "timed_completion_tokens": 50,
         "api_requests": 2,  # 1 + 1
         "successful_requests": 2,  # 1 + 1
         "failed_requests": 0,  # 0 + 0
@@ -216,9 +215,7 @@ async def test_get_aggregated_daily_spend_update_transactions_same_key():
     updates = [{test_key: test_transaction1}, {test_key: test_transaction2}]
 
     # Test aggregation
-    result = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(
-        updates
-    )
+    result = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(updates)
 
     assert len(result) == 1
     assert test_key in result
@@ -235,6 +232,7 @@ async def test_flush_and_get_aggregated_daily_spend_update_transactions(
         "spend": 10.0,
         "prompt_tokens": 100,
         "completion_tokens": 50,
+        "timed_completion_tokens": 50,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -244,6 +242,7 @@ async def test_flush_and_get_aggregated_daily_spend_update_transactions(
         "spend": 5.0,
         "prompt_tokens": 200,
         "completion_tokens": 30,
+        "timed_completion_tokens": 0,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -253,6 +252,7 @@ async def test_flush_and_get_aggregated_daily_spend_update_transactions(
         "spend": 15.0,  # 10 + 5
         "prompt_tokens": 300,  # 100 + 200
         "completion_tokens": 80,  # 50 + 30
+        "timed_completion_tokens": 50,
         "api_requests": 2,  # 1 + 1
         "successful_requests": 2,  # 1 + 1
         "failed_requests": 0,  # 0 + 0
@@ -272,9 +272,7 @@ async def test_flush_and_get_aggregated_daily_spend_update_transactions(
     await daily_spend_update_queue.add_update({test_key: test_transaction2})
 
     # Flush and get aggregated transactions
-    result = (
-        await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
-    )
+    result = await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
 
     assert len(result) == 1
     assert test_key in result
@@ -282,9 +280,7 @@ async def test_flush_and_get_aggregated_daily_spend_update_transactions(
 
 
 @pytest.mark.asyncio
-async def test_queue_max_size_triggers_aggregation(
-    monkeypatch, daily_spend_update_queue
-):
+async def test_queue_max_size_triggers_aggregation(monkeypatch, daily_spend_update_queue):
     """Test that reaching MAX_SIZE_IN_MEMORY_QUEUE triggers aggregation"""
     # Override MAX_SIZE_IN_MEMORY_QUEUE for testing
     litellm._turn_on_debug()
@@ -308,9 +304,7 @@ async def test_queue_max_size_triggers_aggregation(
     assert daily_spend_update_queue.update_queue.qsize() == 1
 
     # Verify the aggregated values
-    result = (
-        await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
-    )
+    result = await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
     assert result[test_key]["spend"] == 6.0
     assert result[test_key]["prompt_tokens"] == 600
     assert result[test_key]["completion_tokens"] == 300
@@ -427,9 +421,7 @@ async def test_cache_token_fields_aggregation(daily_spend_update_queue):
 
 
 @pytest.mark.asyncio
-async def test_queue_size_reduction_with_large_volume(
-    monkeypatch, daily_spend_update_queue
-):
+async def test_queue_size_reduction_with_large_volume(monkeypatch, daily_spend_update_queue):
     """Test that queue size is actually reduced when dealing with many items"""
     # Set a smaller MAX_SIZE for testing
     monkeypatch.setattr(daily_spend_update_queue, "MAX_SIZE_IN_MEMORY_QUEUE", 10)
@@ -470,9 +462,7 @@ async def test_queue_size_reduction_with_large_volume(
     assert daily_spend_update_queue.update_queue.qsize() <= 10
 
     # Verify total costs are correct
-    result = (
-        await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
-    )
+    result = await daily_spend_update_queue.flush_and_get_aggregated_daily_spend_update_transactions()
     print("RESULT", json.dumps(result, indent=4))
 
     assert result[user1_key]["spend"] == 200 * 0.5  # 10.0
@@ -545,6 +535,7 @@ async def test_every_optional_daily_metric_aggregates(daily_spend_update_queue):
     paths, so the driver reads as zero on the dashboard however much it saved.
     """
     test_key = "user1_2023-01-01_key123_claude-haiku-4-5_anthropic"
+
     def _numeric(annotation):
         # additive metrics may be declared NotRequired[float] for rows queued by a pod
         # running the previous release, so unwrap before matching
@@ -584,7 +575,15 @@ async def test_optional_metric_missing_from_an_older_payload_still_aggregates(
 
     await daily_spend_update_queue.add_update({test_key: dict(base)})
     await daily_spend_update_queue.add_update(
-        {test_key: {**base, "autorouter_savings_spend": 0.25, "total_response_time_ms": 900, "timed_requests": 1}}
+        {
+            test_key: {
+                **base,
+                "autorouter_savings_spend": 0.25,
+                "total_response_time_ms": 900,
+                "timed_requests": 1,
+                "timed_completion_tokens": 5,
+            }
+        }
     )
     await daily_spend_update_queue.aggregate_queue_updates()
     updates = await daily_spend_update_queue.flush_all_updates_from_in_memory_queue()
@@ -592,3 +591,22 @@ async def test_optional_metric_missing_from_an_older_payload_still_aggregates(
     assert updates[0][test_key]["autorouter_savings_spend"] == pytest.approx(0.25)
     assert updates[0][test_key]["total_response_time_ms"] == 900
     assert updates[0][test_key]["timed_requests"] == 1
+    assert updates[0][test_key]["timed_completion_tokens"] == 5
+
+
+def test_legacy_timed_request_keeps_timed_tokens_unknown():
+    key = "user1_2023-01-01_key123_gpt-4o_openai"
+    base = {
+        "spend": 1.0,
+        "prompt_tokens": 10,
+        "completion_tokens": 5,
+        "api_requests": 1,
+        "successful_requests": 1,
+        "failed_requests": 0,
+        "timed_requests": 1,
+    }
+    updates = [{key: base}, {key: {**base, "timed_completion_tokens": 5}}]
+
+    aggregated = DailySpendUpdateQueue.get_aggregated_daily_spend_update_transactions(updates)
+
+    assert aggregated[key]["timed_completion_tokens"] is None

tests/unit/proxy/db/test_daily_spend_bulk_upsert.py

@@ -8,6 +8,7 @@ import pytest
 
 from litellm.proxy.db.daily_spend_bulk_upsert import (
     DAILY_SPEND_TABLES,
+    _counter_value,
     build_bulk_upsert,
     conflict_key,
     merge_by_conflict_key,
@@ -34,6 +35,7 @@ def tag_txn(**overrides):
         "endpoint": "/chat/completions",
         "prompt_tokens": 10,
         "completion_tokens": 20,
+        "timed_completion_tokens": 20,
         "spend": 0.25,
         "api_requests": 1,
         "successful_requests": 1,
@@ -71,6 +73,31 @@ def test_null_and_empty_provider_merge_into_one_row(order):
     assert folded["api_requests"] == 4
 
 
+def test_untimed_legacy_rows_do_not_hide_known_timed_tokens_when_rows_merge():
+    legacy = tag_txn()
+    del legacy["timed_completion_tokens"]
+
+    merged = merge_by_conflict_key(TAG_TABLE, (legacy, tag_txn(timed_completion_tokens=7)))
+
+    assert merged[0][1]["timed_completion_tokens"] == 7
+
+
+def test_legacy_timed_requests_keep_timed_tokens_unknown_when_rows_merge():
+    legacy = tag_txn(timed_requests=1)
+    del legacy["timed_completion_tokens"]
+
+    merged = merge_by_conflict_key(TAG_TABLE, (legacy, tag_txn(timed_completion_tokens=7)))
+
+    assert merged[0][1]["timed_completion_tokens"] is None
+
+
+def test_legacy_timed_request_serializes_unknown_token_total():
+    legacy = tag_txn(timed_requests=1)
+    del legacy["timed_completion_tokens"]
+
+    assert _counter_value("timed_completion_tokens", legacy) is None
+
+
 def test_distinct_keys_are_not_merged_and_are_ordered_deterministically():
     unordered = (tag_txn(tag="z-team"), tag_txn(tag="a-team"), tag_txn(tag="m-team"))
 
@@ -87,10 +114,10 @@ def test_one_statement_carries_every_row_in_the_batch():
 
     assert sql.count("INSERT INTO") == 1
     assert len(re.findall(r"ON CONFLICT", sql)) == 1
-    # 25 bound columns per row plus the inlined updated_at, so the row count is what
+    # 26 bound columns per row plus the inlined updated_at, so the row count is what
     # separates one multi-row statement from a hundred single-row ones.
-    assert len(params) == 100 * 25
-    assert "$2500::text" in sql
+    assert len(params) == 100 * 26
+    assert "$2600::text" in sql
     assert sql.count("(NOW() AT TIME ZONE 'UTC')") == 100 + 1
 
 
@@ -125,9 +152,7 @@ def test_counters_increment_rather_than_overwrite(column):
 
 
 def test_request_id_is_preserved_when_a_later_batch_carries_none():
-    sql, params = build_bulk_upsert(
-        TAG_TABLE, merge_by_conflict_key(TAG_TABLE, (tag_txn(request_id=None),))
-    )
+    sql, params = build_bulk_upsert(TAG_TABLE, merge_by_conflict_key(TAG_TABLE, (tag_txn(request_id=None),)))
 
     assert '"request_id" = COALESCE(EXCLUDED."request_id", "LiteLLM_DailyTagSpend"."request_id")' in sql
     assert None in params

tests/unit/proxy/db/test_db_spend_update_writer.py

@@ -103,11 +103,21 @@ async def test_update_database_attributes_router_rejected_failure_to_model_group
     )
 
     with (
-        patch("litellm.proxy.proxy_server.disable_spend_logs", True),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
-        patch("litellm.proxy.proxy_server.prisma_client", MagicMock()),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
-        patch("litellm.proxy.proxy_server.user_api_key_cache", MagicMock()),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
-        patch("litellm.proxy.proxy_server.litellm_proxy_budget_name", "test-budget"),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
-        patch("litellm.proxy.proxy_server.llm_router", llm_router),  # test-quality-ok: get_llm_router reads this proxy_server module global at call time; no injection seam
+        patch(
+            "litellm.proxy.proxy_server.disable_spend_logs", True
+        ),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
+        patch(
+            "litellm.proxy.proxy_server.prisma_client", MagicMock()
+        ),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
+        patch(
+            "litellm.proxy.proxy_server.user_api_key_cache", MagicMock()
+        ),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
+        patch(
+            "litellm.proxy.proxy_server.litellm_proxy_budget_name", "test-budget"
+        ),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
+        patch(
+            "litellm.proxy.proxy_server.llm_router", llm_router
+        ),  # test-quality-ok: get_llm_router reads this proxy_server module global at call time; no injection seam
     ):
         await db_writer.update_database(
             token="test-token",
@@ -320,7 +330,9 @@ async def test_a_routed_request_reaches_the_auto_router_rollup_whether_or_not_sp
     }
 
     with (
-        patch("litellm.proxy.proxy_server.disable_spend_logs", disable_spend_logs),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
+        patch(
+            "litellm.proxy.proxy_server.disable_spend_logs", disable_spend_logs
+        ),  # test-quality-ok: update_database reads this proxy_server module global at call time; no injection seam
         patch("litellm.proxy.proxy_server.prisma_client", prisma),
         patch("litellm.proxy.proxy_server.litellm_proxy_budget_name", "test-budget"),
         patch(
@@ -2912,17 +2924,22 @@ async def test_daily_transaction_carries_compression_saved_tokens():
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("estimate, recorded_savings, expected", [
-    pytest.param(None, None, -0.005, id="plain-classifier-cost"),
-    pytest.param({"version": 1, "status": "unknown"}, None, 0.0, id="unknown"),
-    pytest.param({"version": 2, "status": "unknown"}, None, 0.0, id="unknown-v2"),
-    pytest.param({"version": 1, "status": "unknown"}, -0.003, 0.0, id="unknown-stale-value"),
-    pytest.param({"version": 0, "status": "estimated"}, -0.003, 0.0, id="unsupported-version"),
-    pytest.param({"version": 1, "status": "estimated"}, -0.003, -0.003, id="estimated"),
-    pytest.param(None, -0.003, -0.003, id="legacy"),
-])
+@pytest.mark.parametrize(
+    "estimate, recorded_savings, expected",
+    [
+        pytest.param(None, None, -0.005, id="plain-classifier-cost"),
+        pytest.param({"version": 1, "status": "unknown"}, None, 0.0, id="unknown"),
+        pytest.param({"version": 2, "status": "unknown"}, None, 0.0, id="unknown-v2"),
+        pytest.param({"version": 1, "status": "unknown"}, -0.003, 0.0, id="unknown-stale-value"),
+        pytest.param({"version": 0, "status": "estimated"}, -0.003, 0.0, id="unsupported-version"),
+        pytest.param({"version": 1, "status": "estimated"}, -0.003, -0.003, id="estimated"),
+        pytest.param(None, -0.003, -0.003, id="legacy"),
+    ],
+)
 async def test_daily_transaction_compression_saved_tokens_zero_when_absent(
-    estimate: dict[str, object] | None, recorded_savings: float | None, expected: float,
+    estimate: dict[str, object] | None,
+    recorded_savings: float | None,
+    expected: float,
 ) -> None:
     """Requests without any compression metadata produce a zero count."""
     writer = DBSpendUpdateWriter()
@@ -2941,12 +2958,14 @@ async def test_daily_transaction_compression_saved_tokens_zero_when_absent(
         "prompt_tokens": 100,
         "completion_tokens": 10,
         "spend": 0.01,
-        "metadata": json.dumps({
-            "usage_object": {"prompt_tokens": 100, "completion_tokens": 10},
-            "routing_decision": {"savings_baseline_model": "anthropic/claude-sonnet-5", "classifier_cost": 0.005},
-            "autorouter_savings": recorded_savings,
-            "autorouter_savings_estimate": estimate,
-        }),
+        "metadata": json.dumps(
+            {
+                "usage_object": {"prompt_tokens": 100, "completion_tokens": 10},
+                "routing_decision": {"savings_baseline_model": "anthropic/claude-sonnet-5", "classifier_cost": 0.005},
+                "autorouter_savings": recorded_savings,
+                "autorouter_savings_estimate": estimate,
+            }
+        ),
     }
 
     transaction = await writer._common_add_spend_log_transaction_to_daily_transaction(
@@ -3439,9 +3458,7 @@ async def test_failed_per_entity_increment_from_redis_restores_only_what_may_sti
         )
 
     mock_redis_update_buffer.restore_transactions_to_redis.assert_awaited_once()
-    restored = mock_redis_update_buffer.restore_transactions_to_redis.call_args.kwargs[
-        "db_spend_update_transactions"
-    ]
+    restored = mock_redis_update_buffer.restore_transactions_to_redis.call_args.kwargs["db_spend_update_transactions"]
     assert restored["user_list_transactions"] is None
     assert restored["team_list_transactions"] == {"team-1": 1.5}
     assert restored["key_list_transactions"] == ({"key-1": 1.5} if safe_to_resend else None)
@@ -3758,6 +3775,7 @@ async def test_daily_transaction_rolls_up_response_time_for_successful_requests(
     assert transaction is not None
     assert transaction["total_response_time_ms"] == request_duration_ms
     assert transaction["timed_requests"] == 1
+    assert transaction["timed_completion_tokens"] == 5
 
 
 @pytest.mark.asyncio
@@ -3790,6 +3808,7 @@ async def test_daily_transaction_excludes_untimed_requests_from_response_time(
     assert transaction is not None
     assert transaction["total_response_time_ms"] == 0
     assert transaction["timed_requests"] == 0
+    assert transaction["timed_completion_tokens"] == 0
 
 
 def _deadlock_error():

tests/unit/proxy/management_endpoints/test_common_daily_activity.py

@@ -366,6 +366,7 @@ async def test_get_daily_activity_aggregated_with_endpoint_breakdown():
         "autorouter_savings_spend": 0.0,
         "total_response_time_ms": 0,
         "timed_requests": 0,
+        "timed_completion_tokens": 0,
         "failed_requests": 0,
     }
     mock_rows = [
@@ -950,6 +951,7 @@ def test_update_breakdown_metrics_includes_user_email():
         autorouter_savings_spend=0,
         total_response_time_ms=0,
         timed_requests=0,
+        timed_completion_tokens=0,
         total_tokens=2,
         api_requests=1,
         successful_requests=1,
@@ -1031,6 +1033,7 @@ async def test_tag_daily_activity_metadata_totals_not_zero():
     mock_record_1.autorouter_savings_spend = 0.0
     mock_record_1.total_response_time_ms = 18_000
     mock_record_1.timed_requests = 9
+    mock_record_1.timed_completion_tokens = 200
     mock_record_1.api_requests = 10
     mock_record_1.successful_requests = 9
     mock_record_1.failed_requests = 1
@@ -1057,6 +1060,7 @@ async def test_tag_daily_activity_metadata_totals_not_zero():
     mock_record_2.autorouter_savings_spend = 0.0
     mock_record_2.total_response_time_ms = 2_500
     mock_record_2.timed_requests = 5
+    mock_record_2.timed_completion_tokens = 100
     mock_record_2.api_requests = 5
     mock_record_2.successful_requests = 5
     mock_record_2.failed_requests = 0
@@ -1129,6 +1133,7 @@ async def test_aggregated_activity_preserves_metadata_for_deleted_keys():
         "autorouter_savings_spend": 0.0,
         "total_response_time_ms": 0,
         "timed_requests": 0,
+        "timed_completion_tokens": 0,
         "failed_requests": 0,
     }
     mock_rows = [
@@ -1225,6 +1230,7 @@ async def test_aggregated_activity_flags_only_keys_that_key_info_can_still_resol
         "autorouter_savings_spend": 0.0,
         "total_response_time_ms": 0,
         "timed_requests": 0,
+        "timed_completion_tokens": 0,
         "api_requests": 1,
         "successful_requests": 1,
         "failed_requests": 0,
@@ -1295,6 +1301,7 @@ def _daily_user_spend_record(*, user_id, api_key, spend, model="gpt-4", model_gr
         autorouter_savings_spend=0.0,
         total_response_time_ms=0,
         timed_requests=0,
+        timed_completion_tokens=0,
         api_requests=1,
         successful_requests=1,
         failed_requests=0,
@@ -1444,6 +1451,7 @@ async def test_get_daily_activity_aggregated_empty_result_set():
             "autorouter_savings_spend": None,
             "total_response_time_ms": None,
             "timed_requests": None,
+            "timed_completion_tokens": None,
             "api_requests": None,
             "successful_requests": None,
             "failed_requests": None,
@@ -1492,6 +1500,7 @@ def _no_spend_record():
         autorouter_savings_spend=None,
         total_response_time_ms=None,
         timed_requests=None,
+        timed_completion_tokens=None,
         api_requests=None,
         successful_requests=None,
         failed_requests=None,
@@ -1631,6 +1640,7 @@ def _spend_record(api_key, *, model="gpt-4o-mini-ptu", spend=0.0, ptu_flat_cost=
         autorouter_savings_spend=0,
         total_response_time_ms=0,
         timed_requests=0,
+        timed_completion_tokens=0,
         total_tokens=0,
         api_requests=0,
         successful_requests=0,
@@ -1675,6 +1685,10 @@ def _grouping_row(
     endpoint=None,
     spend=0.0,
     ptu_flat_cost=0.0,
+    completion_tokens=0,
+    timed_completion_tokens=0,
+    total_response_time_ms=0,
+    timed_requests=0,
 ):
     return GroupingSetsRow(
         date="2024-01-01",
@@ -1689,7 +1703,8 @@ def _grouping_row(
         spend=spend,
         ptu_flat_cost=ptu_flat_cost,
         prompt_tokens=0,
-        completion_tokens=0,
+        completion_tokens=completion_tokens,
+        timed_completion_tokens=timed_completion_tokens,
         cache_read_input_tokens=0,
         cache_creation_input_tokens=0,
         compression_saved_tokens=0,
@@ -1697,8 +1712,8 @@ def _grouping_row(
         prompt_caching_savings_spend=0.0,
         gateway_injected_caching_savings_spend=0.0,
         autorouter_savings_spend=0.0,
-        total_response_time_ms=0,
-        timed_requests=0,
+        total_response_time_ms=total_response_time_ms,
+        timed_requests=timed_requests,
         api_requests=0,
         successful_requests=0,
         failed_requests=0,
@@ -1794,6 +1809,108 @@ def test_grouping_sets_dispatcher_populates_every_breakdown_level(ptu_cost_attri
     assert "real-key" in day.breakdown.endpoints["/v1/chat/completions"].api_key_breakdown
 
 
+def test_grouping_sets_dispatcher_returns_provider_throughput_for_models_and_model_groups():
+    from litellm.proxy.management_endpoints.common_daily_activity import (
+        _GROUP_DATE_MODEL_GROUP_PROVIDER,
+        _GROUP_DATE_MODEL_PROVIDER,
+        _aggregate_grouping_sets_records_sync,
+    )
+
+    records = [
+        _grouping_row(
+            _GROUP_DATE_MODEL_PROVIDER,
+            model="gpt-4o",
+            custom_llm_provider="openai",
+            completion_tokens=1900,
+            timed_completion_tokens=900,
+            total_response_time_ms=3000,
+            timed_requests=3,
+        ),
+        _grouping_row(
+            _GROUP_DATE_MODEL_PROVIDER,
+            model="gpt-4o",
+            custom_llm_provider="azure",
+            completion_tokens=400,
+            timed_completion_tokens=400,
+            total_response_time_ms=2000,
+            timed_requests=2,
+        ),
+        _grouping_row(
+            _GROUP_DATE_MODEL_GROUP_PROVIDER,
+            model_group="public-gpt-4o",
+            custom_llm_provider="openai",
+            completion_tokens=900,
+            timed_completion_tokens=900,
+            total_response_time_ms=3000,
+            timed_requests=3,
+        ),
+    ]
+
+    day = _aggregate_grouping_sets_records_sync(records=records, api_key_metadata={})["results"][0]
+
+    assert day.breakdown.models["gpt-4o"].provider_breakdown["openai"].model_dump() == {
+        "timed_completion_tokens": 900,
+        "total_response_time_ms": 3000,
+        "timed_requests": 3,
+        "output_tokens_per_second": 300.0,
+    }
+    assert day.breakdown.models["gpt-4o"].provider_breakdown["azure"].output_tokens_per_second == 200.0
+    assert day.breakdown.model_groups["public-gpt-4o"].provider_breakdown["openai"].output_tokens_per_second == 300.0
+
+
+def test_grouping_sets_dispatcher_returns_zero_for_timed_requests_without_completion_tokens():
+    from litellm.proxy.management_endpoints.common_daily_activity import (
+        _GROUP_DATE_MODEL_PROVIDER,
+        _aggregate_grouping_sets_records_sync,
+    )
+
+    records = [
+        _grouping_row(
+            _GROUP_DATE_MODEL_PROVIDER,
+            model="gpt-4o",
+            completion_tokens=900,
+            timed_completion_tokens=0,
+            total_response_time_ms=3000,
+            timed_requests=1,
+        )
+    ]
+
+    day = _aggregate_grouping_sets_records_sync(records=records, api_key_metadata={})["results"][0]
+
+    assert day.breakdown.models["gpt-4o"].provider_breakdown["openai"].output_tokens_per_second == 0.0
+
+
+def test_grouping_sets_dispatcher_returns_no_throughput_for_legacy_rows_without_timed_tokens():
+    from litellm.proxy.management_endpoints.common_daily_activity import (
+        _GROUP_DATE_MODEL_PROVIDER,
+        _aggregate_grouping_sets_records_sync,
+    )
+
+    records = [
+        _grouping_row(
+            _GROUP_DATE_MODEL_PROVIDER,
+            model="gpt-4o",
+            completion_tokens=900,
+            timed_completion_tokens=None,
+            total_response_time_ms=3000,
+            timed_requests=1,
+        )
+    ]
+
+    day = _aggregate_grouping_sets_records_sync(records=records, api_key_metadata={})["results"][0]
+
+    assert day.breakdown.models["gpt-4o"].provider_breakdown["openai"].output_tokens_per_second is None
+
+
+def test_combined_timed_tokens_propagates_unknown_measurements():
+    from litellm.proxy.management_endpoints.common_daily_activity import _combined_timed_completion_tokens
+    from litellm.types.proxy.management_endpoints.common_daily_activity import ProviderThroughputMetrics
+
+    existing = ProviderThroughputMetrics(timed_completion_tokens=10)
+
+    assert _combined_timed_completion_tokens(existing, None) is None
+
+
 def test_grouping_sets_dispatcher_keeps_ptu_flat_cost_out_of_the_provider_breakdown():
     """Sentinel rows carry no provider, so their flat cost must not surface under the
     "unknown" provider - the per-row path skips them for exactly the same reason."""
@@ -1851,7 +1968,7 @@ def test_update_breakdown_metrics_covers_mcp_endpoint_and_entity(ptu_cost_attrib
         endpoint="/v1/chat/completions",
         spend=5.0,
         prompt_tokens=0,
-        completion_tokens=0,
+        completion_tokens=600,
         cache_read_input_tokens=0,
         cache_creation_input_tokens=0,
         compression_saved_tokens=0,
@@ -1859,8 +1976,9 @@ def test_update_breakdown_metrics_covers_mcp_endpoint_and_entity(ptu_cost_attrib
         prompt_caching_savings_spend=0,
         gateway_injected_caching_savings_spend=0,
         autorouter_savings_spend=0,
-        total_response_time_ms=0,
-        timed_requests=0,
+        total_response_time_ms=2000,
+        timed_requests=2,
+        timed_completion_tokens=600,
         total_tokens=0,
         api_requests=0,
         successful_requests=0,
@@ -1874,6 +1992,8 @@ def test_update_breakdown_metrics_covers_mcp_endpoint_and_entity(ptu_cost_attrib
     assert "real-key" in breakdown.mcp_servers["srv/tool"].api_key_breakdown
     assert "/v1/chat/completions" in breakdown.endpoints
     assert "azure" in breakdown.providers
+    assert breakdown.models["gpt-4o-mini-ptu"].provider_breakdown["azure"].output_tokens_per_second == 300.0
+    assert breakdown.model_groups["grp"].provider_breakdown["azure"].output_tokens_per_second == 300.0
     assert "team-1" in breakdown.entities
     assert "real-key" in breakdown.entities["team-1"].api_key_breakdown
 
@@ -2216,6 +2336,7 @@ async def test_get_daily_activity_aggregated_with_entity_breakdown():
         "autorouter_savings_spend": 0.0,
         "total_response_time_ms": 0,
         "timed_requests": 0,
+        "timed_completion_tokens": 0,
         "failed_requests": 0,
         "prompt_tokens": 0,
         "completion_tokens": 0,

tests/unit/proxy/management_endpoints/test_daily_activity_routes.py

@@ -59,6 +59,7 @@ class _Activity:
     failed_requests: int
     total_response_time_ms: int
     timed_requests: int
+    timed_completion_tokens: int
 
 
 _ENTITY_CASES: Final[tuple[tuple[str, str, str], ...]] = (
@@ -101,6 +102,7 @@ def _activity_for_entity(
             failed_requests=0,
             total_response_time_ms=100,
             timed_requests=1,
+            timed_completion_tokens=5,
         )
         for api_key, date, model, spend, cache_read in key_rows
     )
@@ -157,6 +159,7 @@ def _seeded_activity() -> tuple[_Activity, ...]:
             failed_requests=0,
             total_response_time_ms=100,
             timed_requests=1,
+            timed_completion_tokens=5,
         )
         for table in entity_ids
     )
@@ -180,6 +183,7 @@ def _metrics(rows: Sequence[_Activity]) -> Mapping[str, int | float]:
         "failed_requests": sum(row.failed_requests for row in rows),
         "total_response_time_ms": sum(row.total_response_time_ms for row in rows),
         "timed_requests": sum(row.timed_requests for row in rows),
+        "timed_completion_tokens": sum(row.timed_completion_tokens for row in rows),
     }
 
 

tests/unit/repositories/test_daily_activity_sql.py

@@ -237,6 +237,7 @@ def test_aggregate_query_sums_all_savings_drivers_and_response_time() -> None:
     fields: Final = tuple(field for field in SpendMetrics.model_fields if field.endswith("_savings_spend")) + (
         "total_response_time_ms",
         "timed_requests",
+        "timed_completion_tokens",
     )
 
     assert fields

ui/litellm-dashboard/src/components/UsagePage/dailyActivityApi.ts

@@ -76,6 +76,7 @@ const toMetric = (entry: SchemaMetricWithMetadata): MetricWithMetadata => ({
   api_key_breakdown: Object.fromEntries(
     Object.entries(entry.api_key_breakdown ?? {}).map(([key, value]) => [key, toKeyMetric(value)]),
   ),
+  provider_breakdown: entry.provider_breakdown ?? {},
 });
 
 const toMetricMap = (

ui/litellm-dashboard/src/components/UsagePage/keyActivityData.test.ts

@@ -56,6 +56,14 @@ const aggregatedResponse: DailyActivityAggregatedResponse = {
             metrics: completeMetrics,
             metadata: {},
             api_key_breakdown: { "key-hash": apiKeyActivity },
+            provider_breakdown: {
+              openai: {
+                timed_completion_tokens: 900,
+                output_tokens_per_second: 300,
+                timed_requests: 3,
+                total_response_time_ms: 3000,
+              },
+            },
           },
         },
       },
@@ -105,6 +113,14 @@ describe("key activity data", () => {
         },
       },
     ]);
+    expect(toDailyData(aggregatedResponse)[0].breakdown.models["gpt-4o-mini"].provider_breakdown).toEqual({
+      openai: {
+        timed_completion_tokens: 900,
+        output_tokens_per_second: 300,
+        timed_requests: 3,
+        total_response_time_ms: 3000,
+      },
+    });
   });
 
   it("appends pages without duplicate keys and compares the server offset to the total", () => {

ui/litellm-dashboard/src/components/UsagePage/types.ts

@@ -38,6 +38,14 @@ export interface MetricWithMetadata {
   metrics: SpendMetrics;
   metadata: object;
   api_key_breakdown: { [key: string]: KeyMetricWithMetadata };
+  provider_breakdown?: { [key: string]: ProviderThroughputMetrics };
+}
+
+export interface ProviderThroughputMetrics {
+  timed_completion_tokens?: number | null;
+  total_response_time_ms: number;
+  timed_requests: number;
+  output_tokens_per_second?: number | null;
 }
 
 export interface KeyMetricWithMetadata {
@@ -92,6 +100,7 @@ export interface ModelActivityData {
       cache_creation_input_tokens: number;
       avg_response_time_ms?: number | null;
     };
+    provider_throughput?: Record<string, number | null>;
   }[];
 }
 

ui/litellm-dashboard/src/components/activity_metrics.test.tsx

@@ -1,7 +1,13 @@
 import { fireEvent, render, screen, waitFor } from "@testing-library/react";
 import React from "react";
 import { beforeAll, describe, expect, it, vi } from "vitest";
-import { ActivityMetrics, formatKeyLabel, processActivityData, ResponseTimeTooltip } from "./activity_metrics";
+import {
+  ActivityMetrics,
+  formatKeyLabel,
+  processActivityData,
+  providerThroughputChartData,
+  ResponseTimeTooltip,
+} from "./activity_metrics";
 import type { ChartTooltipProps } from "@/components/shared/charts";
 import { Team } from "./key_team_helpers/key_list";
 import { DailyData, KeyMetricWithMetadata, ModelActivityData } from "./UsagePage/types";
@@ -1402,6 +1408,38 @@ describe("processActivityData", () => {
     expect(result["gpt-5.5"].total_timed_requests).toBe(0);
     expect(result["gpt-5.5"].daily_data[0].metrics.avg_response_time_ms).toBeNull();
   });
+
+  it("preserves daily provider throughput for model and model-group views", () => {
+    const metric = {
+      metrics: EMPTY_SPEND_METRICS,
+      metadata: {},
+      api_key_breakdown: {},
+      provider_breakdown: {
+        openai: {
+          timed_completion_tokens: 900,
+          total_response_time_ms: 3000,
+          timed_requests: 3,
+          output_tokens_per_second: 300,
+        },
+      },
+    };
+    const activity: { results: DailyData[] } = {
+      results: [
+        createMockDailyData("2025-01-01", EMPTY_SPEND_METRICS, {
+          ...EMPTY_BREAKDOWN,
+          models: { "gpt-4o": metric },
+          model_groups: { "public-gpt-4o": metric },
+        }),
+      ],
+    };
+
+    expect(processActivityData(activity, "models")["gpt-4o"].daily_data[0].provider_throughput).toEqual({
+      openai: 300,
+    });
+    expect(processActivityData(activity, "model_groups")["public-gpt-4o"].daily_data[0].provider_throughput).toEqual({
+      openai: 300,
+    });
+  });
 });
 
 describe("ActivityMetrics response time", () => {
@@ -1482,6 +1520,58 @@ describe("ActivityMetrics response time", () => {
   });
 });
 
+describe("ActivityMetrics provider throughput", () => {
+  const model = createMockModelActivityData("GPT-4o", {
+    daily_data: [
+      {
+        ...createMockModelActivityData("GPT-4o").daily_data[0],
+        date: "2025-01-01",
+        provider_throughput: { openai: 300, azure: 200 },
+      },
+      {
+        ...createMockModelActivityData("GPT-4o").daily_data[0],
+        date: "2025-01-02",
+        provider_throughput: { openai: 250 },
+      },
+    ],
+  });
+
+  it("renders one daily line per provider", () => {
+    render(<ActivityMetrics modelMetrics={{ "gpt-4o": model }} />);
+
+    expect(screen.getByText("Output tokens per second of response time")).toBeInTheDocument();
+    expect(screen.getByText("Openai")).toBeInTheDocument();
+    expect(screen.getByText("Azure")).toBeInTheDocument();
+    expect(screen.getAllByText("2025-01-01").length).toBeGreaterThan(0);
+    expect(screen.getAllByText("2025-01-02").length).toBeGreaterThan(0);
+  });
+
+  it("keeps missing provider dates as gaps", () => {
+    expect(providerThroughputChartData(model.daily_data)).toEqual({
+      providers: ["azure", "openai"],
+      data: [
+        { date: "2025-01-01", azure: 200, openai: 300 },
+        { date: "2025-01-02", azure: null, openai: 250 },
+      ],
+    });
+  });
+
+  it("hides the chart when every throughput value is unavailable", () => {
+    const unavailable = createMockModelActivityData("GPT-4o", {
+      daily_data: [
+        {
+          ...createMockModelActivityData("GPT-4o").daily_data[0],
+          provider_throughput: { openai: null },
+        },
+      ],
+    });
+
+    render(<ActivityMetrics modelMetrics={{ "gpt-4o": unavailable }} />);
+
+    expect(screen.queryByText("Output tokens per second of response time")).not.toBeInTheDocument();
+  });
+});
+
 describe("formatKeyLabel", () => {
   it("should return key_alias when no team_id is present", () => {
     const modelData = createMockKeyMetricWithMetadata({

ui/litellm-dashboard/src/components/activity_metrics.tsx

@@ -4,6 +4,7 @@ import {
   type ChartTooltipProps,
   CustomLegend,
   CustomTooltip,
+  DEFAULT_COLOR_CYCLE,
   formatCategoryName,
   LineChart,
   ValueTooltip,
@@ -18,7 +19,13 @@ import { Team } from "./key_team_helpers/key_list";
 import KeyModelUsageView from "./UsagePage/components/KeyModelUsageView";
 import { keyActivityLabel } from "./UsagePage/keyActivityLabel";
 import type { ModelTopKeysResponse } from "./UsagePage/dailyActivityApi";
-import { DailyData, KeyMetricWithMetadata, ModelActivityData, TopModelData } from "./UsagePage/types";
+import {
+  DailyData,
+  KeyMetricWithMetadata,
+  MetricWithMetadata,
+  ModelActivityData,
+  TopModelData,
+} from "./UsagePage/types";
 import { averageResponseTimeMs, formatResponseTime, valueFormatter } from "./UsagePage/utils/value_formatters";
 
 interface ActivityMetricsProps {
@@ -41,6 +48,30 @@ export const ResponseTimeTooltip = ({ active, payload, label }: ChartTooltipProp
   />
 );
 
+const formatTokensPerSecond = (value: number): string =>
+  `${value.toLocaleString(undefined, { maximumFractionDigits: 2 })} tokens/s`;
+
+export const providerThroughputChartData = (dailyData: ModelActivityData["daily_data"]) => {
+  const providers = Array.from(new Set(dailyData.flatMap((day) => Object.keys(day.provider_throughput ?? {}))))
+    .filter((provider) =>
+      dailyData.some((day) => {
+        const value = day.provider_throughput?.[provider];
+        return typeof value === "number" && Number.isFinite(value);
+      }),
+    )
+    .sort();
+  const data = dailyData.map((day) => ({
+    date: day.date,
+    ...Object.fromEntries(
+      providers.map((provider) => {
+        const value = day.provider_throughput?.[provider];
+        return [provider, typeof value === "number" && Number.isFinite(value) ? value : null];
+      }),
+    ),
+  }));
+  return { providers, data };
+};
+
 const ModelTopKeys = ({
   modelName,
   fetchTopApiKeys,
@@ -178,6 +209,8 @@ export const ModelSection = ({
   hidePromptCachingMetrics?: boolean;
   fetchTopApiKeys?: (model: string) => Promise<ModelTopKeysResponse>;
 }) => {
+  const throughputChart = providerThroughputChartData(metrics.daily_data);
+
   return (
     <div className="space-y-2">
       {/* Summary Cards */}
@@ -316,6 +349,27 @@ export const ModelSection = ({
           </Card>
         )}
 
+        {throughputChart.providers.length > 0 && (
+          <Card>
+            <CardContent>
+              <div className="flex justify-between items-center">
+                <h3 className="text-lg font-medium text-foreground">Output tokens per second of response time</h3>
+                <CustomLegend categories={throughputChart.providers} colors={DEFAULT_COLOR_CYCLE} />
+              </div>
+              <LineChart
+                className="mt-4"
+                data={throughputChart.data}
+                index="date"
+                categories={throughputChart.providers}
+                colors={DEFAULT_COLOR_CYCLE}
+                valueFormatter={formatTokensPerSecond}
+                connectNulls={false}
+                showLegend={false}
+              />
+            </CardContent>
+          </Card>
+        )}
+
         <Card>
           <CardContent>
             <div className="flex justify-between items-center">
@@ -683,6 +737,13 @@ export const processActivityData = (
       modelMetrics[model].total_response_time_ms =
         (modelMetrics[model].total_response_time_ms ?? 0) + dayResponseTimeMs;
       modelMetrics[model].total_timed_requests = (modelMetrics[model].total_timed_requests ?? 0) + dayTimedRequests;
+      const providerBreakdown = (modelData as MetricWithMetadata).provider_breakdown ?? {};
+      const providerThroughput = Object.fromEntries(
+        Object.entries(providerBreakdown).map(([provider, providerMetrics]) => [
+          provider,
+          providerMetrics.output_tokens_per_second ?? null,
+        ]),
+      );
 
       // Add daily data
       modelMetrics[model].daily_data.push({
@@ -699,6 +760,7 @@ export const processActivityData = (
           cache_creation_input_tokens: modelData.metrics.cache_creation_input_tokens || 0,
           avg_response_time_ms: averageResponseTimeMs(dayResponseTimeMs, dayTimedRequests),
         },
+        ...(Object.keys(providerThroughput).length > 0 ? { provider_throughput: providerThroughput } : {}),
       });
     });
   });

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -38122,6 +38122,10 @@ export interface components {
                 [key: string]: unknown;
             };
             metrics: components["schemas"]["SpendMetrics"];
+            /** Provider Breakdown */
+            provider_breakdown?: {
+                [key: string]: components["schemas"]["ProviderThroughputMetrics"];
+            };
         };
         /** Mode */
         Mode: {
@@ -41755,6 +41759,23 @@ export interface components {
             /** Tooltip */
             tooltip?: string | null;
         };
+        /** ProviderThroughputMetrics */
+        ProviderThroughputMetrics: {
+            /** Output Tokens Per Second */
+            output_tokens_per_second?: number | null;
+            /** Timed Completion Tokens */
+            timed_completion_tokens?: number | null;
+            /**
+             * Timed Requests
+             * @default 0
+             */
+            timed_requests: number;
+            /**
+             * Total Response Time Ms
+             * @default 0
+             */
+            total_response_time_ms: number;
+        };
         /**
          * ProxyChatCompletionRequest
          * @description Pydantic model for chat completion requests that includes both OpenAI standard fields