Merge 7c7af51185 into 72a9e1cf37

tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py

@@ -59,17 +59,17 @@ class GptProfile(NamedTuple):
 GPT_5_6_PROFILES = [
     GptProfile(
         model_id="us.openai.gpt-5.6-sol",
-        input_cost=5.5e-06, input_cost_above_272k=1.1e-05,
-        cache_write=6.875e-06, cache_write_above_272k=1.375e-05,
-        cache_read=5.5e-07, cache_read_above_272k=1.1e-06,
-        output_cost=3.3e-05, output_cost_above_272k=4.95e-05,
+        input_cost=4.4e-06, input_cost_above_272k=8.8e-06,
+        cache_write=5.5e-06, cache_write_above_272k=1.1e-05,
+        cache_read=4.4e-07, cache_read_above_272k=8.8e-07,
+        output_cost=2.2e-05, output_cost_above_272k=3.3e-05,
     ),
     GptProfile(
         model_id="global.openai.gpt-5.6-sol",
-        input_cost=5e-06, input_cost_above_272k=1e-05,
-        cache_write=6.25e-06, cache_write_above_272k=1.25e-05,
-        cache_read=5e-07, cache_read_above_272k=1e-06,
-        output_cost=3e-05, output_cost_above_272k=4.5e-05,
+        input_cost=4e-06, input_cost_above_272k=8e-06,
+        cache_write=5e-06, cache_write_above_272k=1e-05,
+        cache_read=4e-07, cache_read_above_272k=8e-07,
+        output_cost=2e-05, output_cost_above_272k=3e-05,
     ),
     GptProfile(
         model_id="us.openai.gpt-5.6-terra",
@@ -221,7 +221,7 @@ def test_bedrock_gpt_5_6_above_272k_tier_applies_to_cost(local_model_cost_map):
         custom_llm_provider="bedrock",
     )
 
-    assert cost == pytest.approx((300000 * 1.1e-05) + (1000 * 4.95e-05), rel=1e-9)
+    assert cost == pytest.approx((300000 * 8.8e-06) + (1000 * 3.3e-05), rel=1e-9)
 
 
 def test_bedrock_gpt_5_6_bills_cache_read_tokens(local_model_cost_map):
@@ -241,10 +241,10 @@ def test_bedrock_gpt_5_6_bills_cache_read_tokens(local_model_cost_map):
         custom_llm_provider="bedrock",
     )
 
-    expected = (2 * 5.5e-06) + (15609 * 5.5e-07) + (5 * 3.3e-05)
+    expected = (2 * 4.4e-06) + (15609 * 4.4e-07) + (5 * 2.2e-05)
     assert cost == pytest.approx(expected, rel=1e-9)
     # Without cache_read_input_token_cost the cached prefix bills at zero.
-    assert cost > (15611 * 5.5e-06) * 0.1
+    assert cost > (15611 * 4.4e-06) * 0.1
 
 
 def test_bedrock_gpt_5_6_bills_cache_write_tokens(local_model_cost_map):
@@ -263,7 +263,7 @@ def test_bedrock_gpt_5_6_bills_cache_write_tokens(local_model_cost_map):
         custom_llm_provider="bedrock",
     )
 
-    expected = (2 * 5.5e-06) + (15609 * 6.875e-06) + (5 * 3.3e-05)
+    expected = (2 * 4.4e-06) + (15609 * 5.5e-06) + (5 * 2.2e-05)
     assert cost == pytest.approx(expected, rel=1e-9)
 
 

tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py

@@ -1683,10 +1683,19 @@ class TestBedrockMantleResponsesPricing:
         assert info["cache_read_input_token_cost"] == pytest.approx(2.75e-07)
         assert info["max_input_tokens"] == 1050000
 
+    def test_gpt_5_6_cyber_pricing_and_mode(self, local_cost_map):
+        info = litellm.get_model_info("bedrock_mantle/openai.gpt-5.6-cyber")
+        assert info["mode"] == "responses"
+        assert info["input_cost_per_token"] == pytest.approx(1.375e-05)
+        assert info["cache_creation_input_token_cost"] == pytest.approx(1.71875e-05)
+        assert info["cache_read_input_token_cost"] == pytest.approx(1.375e-06)
+        assert info["output_cost_per_token"] == pytest.approx(8.25e-05)
+        assert info["max_input_tokens"] == 272000
+
     @pytest.mark.parametrize(
         "model, input_cost, cache_creation_cost, cache_read_cost, output_cost",
         [
-            ("openai.gpt-5.6-sol", 5.5e-06, 6.875e-06, 5.5e-07, 3.3e-05),
+            ("openai.gpt-5.6-sol", 4.4e-06, 5.5e-06, 4.4e-07, 2.2e-05),
             ("openai.gpt-5.6-terra", 2.2e-06, 2.75e-06, 2.2e-07, 1.32e-05),
             ("openai.gpt-5.6-luna", 2.2e-07, 2.75e-07, 2.2e-08, 1.32e-06),
         ],
@@ -1709,7 +1718,7 @@ class TestBedrockMantleResponsesPricing:
     @pytest.mark.parametrize(
         "model, input_cost, output_cost",
         [
-            ("openai.gpt-5.6-sol", 5.5e-06, 3.3e-05),
+            ("openai.gpt-5.6-sol", 4.4e-06, 2.2e-05),
             ("openai.gpt-5.6-terra", 2.2e-06, 1.32e-05),
             ("openai.gpt-5.6-luna", 2.2e-07, 1.32e-06),
         ],

tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py

@@ -87,3 +87,56 @@ def test_anthropic_sonnet_1hr_cache_write_pricing(
         ), f"{model_key}: long-context 1hr/5min ratio is {ratio_lc}, expected 1.6"
     else:
         assert "cache_creation_input_token_cost_above_1hr_above_200k_tokens" not in info
+
+
+CLAUDE_3_EXPECTED = [
+    ("claude-3-haiku-20240307", 5e-07),
+    ("claude-3-opus-20240229", 3e-05),
+]
+
+
+@pytest.mark.parametrize("model_key, expected_1hr", CLAUDE_3_EXPECTED)
+def test_claude_3_1hr_cache_write_pricing(model_data, model_key, expected_1hr):
+    """Haiku 3 and Opus 3 both carried Sonnet's 6e-06 1hr rate, overbilling Haiku 3
+    1-hour cache writes 12x and underbilling Opus 3 5x."""
+    info = model_data[model_key]
+
+    assert info["cache_creation_input_token_cost_above_1hr"] == expected_1hr
+
+
+@pytest.mark.parametrize("model_key, expected_1hr", CLAUDE_3_EXPECTED)
+def test_backup_matches_main_for_claude_3_1hr_cache_write(model_key, expected_1hr):
+    json_path = os.path.join(
+        os.path.dirname(__file__),
+        "../../litellm/model_prices_and_context_window_backup.json",
+    )
+    with open(json_path) as f:
+        backup = json.load(f)
+
+    assert (
+        backup[model_key]["cache_creation_input_token_cost_above_1hr"] == expected_1hr
+    )
+
+
+def test_first_party_anthropic_1hr_cache_writes_are_2x_base_input(model_data):
+    """Anthropic charges 1-hour cache writes at 2x base input for every first-party
+    model, so any entry that drifts off that multiple is a copy-paste error."""
+    offenders = tuple(
+        (
+            model_key,
+            info["input_cost_per_token"],
+            info["cache_creation_input_token_cost_above_1hr"],
+        )
+        for model_key, info in model_data.items()
+        if isinstance(info, dict)
+        and info.get("litellm_provider") == "anthropic"
+        and info.get("input_cost_per_token")
+        and info.get("cache_creation_input_token_cost_above_1hr")
+        and abs(
+            info["cache_creation_input_token_cost_above_1hr"]
+            - 2 * info["input_cost_per_token"]
+        )
+        > 1e-12
+    )
+
+    assert offenders == (), f"1hr cache write is not 2x base input for: {offenders}"

tests/test_litellm/test_fireworks_serverless_model_costs.py

@@ -0,0 +1,86 @@
+"""
+Validate the Fireworks AI Serverless entry added for #37274 exists in
+`model_prices_and_context_window.json` and that the bare Fireworks model ID
+resolves through `get_model_info`.
+
+Pricing as published at https://docs.fireworks.ai/serverless/pricing
+(USD per 1M tokens, uncached input / cached input / output):
+
+  accounts/fireworks/models/deepseek-v4-pro-0813  -> $1.32 / $0.044 / $3.96
+"""
+
+import json
+import os
+
+import pytest
+
+import litellm
+from litellm.utils import get_model_info
+
+
+@pytest.fixture(scope="module", autouse=True)
+def _local_model_cost_map():
+    """
+    Point litellm at the bundled cost map for the duration of this module
+    only. ``mp.undo()`` restores both the environment variable and
+    ``litellm.model_cost`` so nothing leaks into later tests.
+    """
+    mp = pytest.MonkeyPatch()
+    mp.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    mp.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    get_model_info.cache_clear()
+    yield
+    mp.undo()
+    get_model_info.cache_clear()
+
+
+NEW_ENTRIES = {
+    "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813": {
+        "input_cost_per_token": 1.32e-06,
+        "cache_read_input_token_cost": 4.4e-08,
+        "output_cost_per_token": 3.96e-06,
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 131072,
+    },
+}
+
+
+@pytest.fixture(scope="module")
+def model_data():
+    json_path = os.path.join(
+        os.path.dirname(__file__), "../../model_prices_and_context_window.json"
+    )
+    with open(json_path) as f:
+        return json.load(f)
+
+
+def test_fireworks_serverless_entries_exist(model_data):
+    """The new prefixed entry carries the pricing and metadata from #37274."""
+    for key, expected in NEW_ENTRIES.items():
+        assert key in model_data, f"{key} is missing from model_prices_and_context_window.json"
+        entry = model_data[key]
+        for field, value in expected.items():
+            assert entry[field] == pytest.approx(value), f"{key}.{field}"
+        assert entry["litellm_provider"] == "fireworks_ai"
+        assert entry["mode"] == "chat"
+        assert entry["supports_function_calling"] is True
+        assert entry["supports_vision"] is False
+
+
+def test_bare_fireworks_ids_resolve_through_prefixed_entries():
+    """Bare IDs from #37274 resolve via the provider-prefix lookup path."""
+    for bare_id, prefixed_key in [
+        (
+            "accounts/fireworks/models/deepseek-v4-pro-0813",
+            "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813",
+        ),
+    ]:
+        info = get_model_info(model=bare_id, custom_llm_provider="fireworks_ai")
+        expected = NEW_ENTRIES[prefixed_key]
+        assert info.get("key") == prefixed_key
+        assert info["litellm_provider"] == "fireworks_ai"
+        assert info["input_cost_per_token"] == pytest.approx(expected["input_cost_per_token"])
+        assert info["cache_read_input_token_cost"] == pytest.approx(expected["cache_read_input_token_cost"])
+        assert info["output_cost_per_token"] == pytest.approx(expected["output_cost_per_token"])
+        assert info["max_input_tokens"] == expected["max_input_tokens"]
+        assert info["max_output_tokens"] == expected["max_output_tokens"]

tests/test_litellm/test_together_ai_model_metadata.py

@@ -31,16 +31,16 @@ SERVERLESS_CHAT_MODELS: Final = (
     "together_ai/meta-models/Muse-Glimmer-30B",
     "together_ai/google/gemma-4-31B-it",
     "together_ai/pearl-ai/gemma-4-31b-it",
-    "together_ai/google/gemma-3n-E4B-it",
     "together_ai/arize-ai/qwen-2-1.5b-instruct",
     "together_ai/Prism-ML/Ternary-Bonsai-27B",
-    "together_ai/meta-llama/Llama-Guard-4-12B",
     "together_ai/openai/gpt-oss-120b",
     "together_ai/openai/gpt-oss-20b",
     "together_ai/meta-llama/Llama-3.3-70B-Instruct-Turbo",
 )
 
 DEPRECATED_MODELS: Final = {
+    "together_ai/google/gemma-3n-E4B-it": "2026-08-25",
+    "together_ai/meta-llama/Llama-Guard-4-12B": "2026-08-25",
     "together_ai/Qwen/Qwen3-235B-A22B-Instruct-2507-tput": "2026-07-10",
     "together_ai/Qwen/Qwen3.5-397B-A17B": "2026-06-29",
     "together_ai/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8": "2026-06-04",