fix(model_prices): verified registry audit, Databricks Sep-2026 catalog, realtime image pricing, deprecation dates

model_prices_and_context_window.schema.json

@@ -79,6 +79,11 @@
           "minimum": 0,
           "description": "USD per token written to the provider's prompt cache."
         },
+        "cache_creation_input_token_cost_above_128k_tokens": {
+          "type": "number",
+          "minimum": 0,
+          "description": "Rate applied once the prompt exceeds the token threshold in the field name."
+        },
         "cache_creation_input_token_cost_above_1hr": {
           "type": "number",
           "minimum": 0,
@@ -94,6 +99,11 @@
           "minimum": 0,
           "description": "Rate applied once the prompt exceeds the token threshold in the field name."
         },
+        "cache_creation_input_token_cost_above_256k_tokens": {
+          "type": "number",
+          "minimum": 0,
+          "description": "Rate applied once the prompt exceeds the token threshold in the field name."
+        },
         "cache_creation_input_token_cost_above_272k_tokens": {
           "type": "number",
           "minimum": 0,
@@ -128,6 +138,11 @@
           "minimum": 0,
           "description": "USD per prompt token served from the provider's prompt cache."
         },
+        "cache_read_input_token_cost_above_128k_tokens": {
+          "type": "number",
+          "minimum": 0,
+          "description": "Rate applied once the prompt exceeds the token threshold in the field name."
+        },
         "cache_read_input_token_cost_above_200k_tokens": {
           "type": "number",
           "minimum": 0,
@@ -138,6 +153,11 @@
           "minimum": 0,
           "description": "Priority service-tier rate for the same-named base field."
         },
+        "cache_read_input_token_cost_above_256k_tokens": {
+          "type": "number",
+          "minimum": 0,
+          "description": "Rate applied once the prompt exceeds the token threshold in the field name."
+        },
         "cache_read_input_token_cost_above_272k_tokens": {
           "type": "number",
           "minimum": 0,

tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py

@@ -4630,6 +4630,28 @@ def test_generic_cost_per_token_grok_46_long_context(_local_model_cost_map):
     assert completion_cost == pytest.approx(1_000 * 1.2e-05)
 
 
+@pytest.mark.parametrize(
+    ("model", "provider", "image_token_rate"),
+    [
+        ("gpt-realtime-2.1", "openai", 5e-06),
+        ("gpt-realtime-2.1-mini", "openai", 8e-07),
+        ("azure/gpt-realtime-2.1", "azure", 5e-06),
+        ("azure/gpt-realtime-2.1-mini", "azure", 8e-07),
+    ],
+)
+def test_realtime_image_tokens_priced_per_token(model, provider, image_token_rate, _local_model_cost_map):
+    """Realtime image input is billed per 1M image tokens, not per image."""
+    usage = Usage(
+        prompt_tokens=1_100,
+        completion_tokens=0,
+        total_tokens=1_100,
+        prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=100, image_tokens=1_000),
+    )
+    prompt_cost, _ = generic_cost_per_token(model=model, usage=usage, custom_llm_provider=provider)
+    text_rate = litellm.model_cost[model]["input_cost_per_token"]
+    assert prompt_cost == pytest.approx(100 * text_rate + 1_000 * image_token_rate)
+
+
 @pytest.mark.parametrize(
     ("response_quality", "requested_quality", "expected_cost"),
     [

tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py

@@ -9,7 +9,6 @@
 
 import pytest
 
-
 from litellm.litellm_core_utils.fallback_generalizations import (
     get_fallback_generalization_rules,
     match_capability_generalizations,
@@ -248,6 +247,57 @@ def test_azure_ai_claude_1m_context_entries(cost_map: dict):
         assert cost_map[model]["max_input_tokens"] == 200000, model
 
 
+# OpenRouter headline rates from GET https://openrouter.ai/api/v1/models.
+# These were the catalog values that disagreed with that API (and, for the
+# two spotlight models, the public model pages that their source fields cite).
+_OPENROUTER_LIVE_COSTS = {
+    "openrouter/qwen/qwen3.5-plus-02-15": (2.6e-07, 1.56e-06, None),
+    "openrouter/openai/gpt-oss-120b": (3.7e-08, 1.7e-07, None),
+    "openrouter/qwen/qwen3-coder-plus": (6.5e-07, 3.25e-06, None),
+    "openrouter/qwen/qwen3.5-flash-02-23": (6.5e-08, 2.6e-07, None),
+    "openrouter/qwen/qwen3.5-27b": (1.95e-07, 1.56e-06, None),
+    "openrouter/gryphe/mythomax-l2-13b": (6e-08, 6e-08, None),
+    "openrouter/mancer/weaver": (4e-07, 7.5e-07, None),
+    "openrouter/xiaomi/mimo-v2.5-pro": (4.35e-07, 8.7e-07, 3.6e-09),
+    "openrouter/moonshotai/kimi-k2.5": (4.5e-07, 2.25e-06, 7e-08),
+    "openrouter/z-ai/glm-5": (6e-07, 1.92e-06, None),
+}
+
+_OPENROUTER_STALE_COSTS = {
+    "openrouter/qwen/qwen3.5-plus-02-15": (4e-07, 2.4e-06),
+    "openrouter/openai/gpt-oss-120b": (1.8e-07, 8e-07),
+    "openrouter/gryphe/mythomax-l2-13b": (1.875e-06, 1.875e-06),
+}
+
+
+@pytest.mark.parametrize(
+    "cost_map",
+    [_load_root_cost_map(), GetModelCostMap.load_local_model_cost_map()],
+    ids=["root", "bundled_backup"],
+)
+def test_openrouter_catalog_costs_match_live_headline_rates(cost_map: dict):
+    """openrouter/* spend tracking reads these catalog fields. The values must
+    stay aligned with OpenRouter's published headline rate, not the stale
+    figures that over/under-counted by up to 30x. Both maps are checked so
+    the root file and bundled backup cannot drift apart."""
+    control = cost_map["openrouter/anthropic/claude-opus-5"]
+    assert control["input_cost_per_token"] == 5e-06
+    assert control["output_cost_per_token"] == 2.5e-05
+    assert control["cache_read_input_token_cost"] == 5e-07
+
+    for model, (inp, out, cache) in _OPENROUTER_LIVE_COSTS.items():
+        entry = cost_map[model]
+        assert entry["input_cost_per_token"] == inp, model
+        assert entry["output_cost_per_token"] == out, model
+        if cache is not None:
+            assert entry["cache_read_input_token_cost"] == cache, model
+
+    for model, (stale_in, stale_out) in _OPENROUTER_STALE_COSTS.items():
+        entry = cost_map[model]
+        assert entry["input_cost_per_token"] != stale_in, model
+        assert entry["output_cost_per_token"] != stale_out, model
+
+
 def test_get_model_cost_map_stamps_loaded_at(monkeypatch):
     """The load time feeds each pod's reload-due decision; a load that does not stamp it
     would make manual reload requests race the proxy's startup"""

tests/test_litellm/llms/azure_ai/test_azure_ai_cost_calculator.py

@@ -499,3 +499,32 @@ def test_service_tier_flex_lower_cost(self):
 
         assert flex_prompt < standard_prompt
         assert flex_completion < standard_completion
+
+
+def test_codestral_2501_model_info_and_cost(local_model_cost_map):
+    model_info = get_model_info(model="Codestral-2501", custom_llm_provider="azure_ai")
+    usage = Usage(prompt_tokens=1_000_000, completion_tokens=1_000_000, total_tokens=2_000_000)
+
+    prompt_cost, completion_cost = cost_per_token(model="Codestral-2501", usage=usage)
+
+    assert model_info["mode"] == "chat"
+    assert model_info["max_input_tokens"] == 256000
+    assert model_info["max_output_tokens"] == 4096
+    assert prompt_cost == pytest.approx(0.3)
+    assert completion_cost == pytest.approx(0.9)
+
+
+def test_mai_thinking_1_model_info_and_cost(local_model_cost_map):
+    model_info = get_model_info(model="MAI-Thinking-1", custom_llm_provider="azure_ai")
+    usage = Usage(prompt_tokens=1_000_000, completion_tokens=1_000_000, total_tokens=2_000_000)
+
+    prompt_cost, completion_cost = cost_per_token(model="MAI-Thinking-1", usage=usage)
+
+    assert model_info["mode"] == "chat"
+    assert model_info["max_input_tokens"] == 256000
+    assert model_info["max_output_tokens"] == 64000
+    assert model_info["cache_read_input_token_cost"] == pytest.approx(2e-07)
+    assert model_info["supports_reasoning"] is True
+    assert model_info["supports_function_calling"] is True
+    assert prompt_cost == pytest.approx(2.0)
+    assert completion_cost == pytest.approx(8.0)

tests/test_litellm/llms/azure_ai/test_azure_ai_fw_models_metadata.py

@@ -176,6 +176,7 @@ def test_azure_ai_fw_model_info(use_local_model_cost_map, model_key, expected):
         ("FW-MiniMax-M2.5", 0.33, 1.32),
         ("FW-Inkling", 1.0, 4.05),
         ("FW-Nemotron-3-Ultra-NVFP4", 0.6, 2.4),
+        ("FW-Nemotron-Lightning-3.5-30B-A3B", 0.06, 0.22),
     ],
 )
 def test_azure_ai_fw_cost_per_token(
@@ -196,6 +197,30 @@ def test_azure_ai_fw_cost_per_token(
     assert completion_cost == pytest.approx(expected_completion)
 
 
+def test_azure_ai_fw_nemotron_lightning_model_info(use_local_model_cost_map):
+    model_info = use_local_model_cost_map.get_model_info(model="azure_ai/FW-Nemotron-Lightning-3.5-30B-A3B")
+
+    assert model_info["litellm_provider"] == "azure_ai"
+    assert model_info["mode"] == "chat"
+    assert model_info["input_cost_per_token"] == pytest.approx(6e-08)
+    assert model_info["output_cost_per_token"] == pytest.approx(2.2e-07)
+    assert model_info["cache_read_input_token_cost"] == pytest.approx(1e-08)
+    assert model_info["max_input_tokens"] == 262144
+    assert model_info["supports_function_calling"] is True
+    assert model_info["supports_reasoning"] is True
+    assert model_info["supports_tool_choice"] is True
+    assert model_info["supports_prompt_caching"] is True
+    assert model_info["supports_vision"] is False
+
+
+def test_azure_ai_fw_nemotron_lightning_supports_tool_choice(use_local_model_cost_map):
+    from litellm.llms.azure_ai.chat.transformation import AzureAIStudioConfig
+
+    supported_params = AzureAIStudioConfig().get_supported_openai_params("FW-Nemotron-Lightning-3.5-30B-A3B")
+
+    assert "tool_choice" in supported_params
+
+
 def test_azure_ai_fw_kimi_k26_case_insensitive_lookup(use_local_model_cost_map):
     upper = use_local_model_cost_map.get_model_info(model="azure_ai/FW-Kimi-K2.6")
     lower = use_local_model_cost_map.get_model_info(model="azure_ai/fw-kimi-k2.6")

tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py

@@ -18,6 +18,10 @@
     "databricks/databricks-claude-opus-5",
     "databricks/databricks-claude-sonnet-5",
     "databricks/databricks-claude-fable-5",
+    "databricks/databricks-claude-fable-5-1",
+    "databricks/databricks-gpt-5-6-sol",
+    "databricks/databricks-gpt-5-6-terra",
+    "databricks/databricks-gpt-5-6-luna",
 )
 
 DOLLARS_PER_DBU: Final = Decimal("0.070")
@@ -28,6 +32,7 @@
     "cache_read_input_token_cost",
 )
 PUBLISHED_DBU_PER_MILLION: Final = {
+    "databricks/databricks-claude-fable-5-1": ("142.858", "714.286", "178.572", "3.572"),
     "databricks/databricks-claude-fable-5": ("142.858", "714.286", "178.572", "14.286"),
     "databricks/databricks-claude-opus-5": ("71.429", "357.143", "89.286", "7.143"),
     "databricks/databricks-claude-opus-4-8": ("71.429", "357.143", "89.286", "7.143"),
@@ -52,9 +57,17 @@
     "databricks/databricks-gpt-5-2": ("25.000", "200.000", "25.000", "2.500"),
     "databricks/databricks-gpt-5-2-codex": ("25.000", "200.000", "25.000", "2.500"),
     "databricks/databricks-gpt-5-3-codex": ("25.000", "200.000", "25.000", "2.500"),
+    "databricks/databricks-gpt-5-6-sol": ("57.143", "285.714", "71.429", "5.714"),
+    "databricks/databricks-gpt-5-6-terra": ("35.714", "214.286", "44.643", "3.571"),
+    "databricks/databricks-gpt-5-6-luna": ("14.286", "85.714", "17.857", "1.429"),
+    "databricks/databricks-gpt-5-5": ("71.429", "428.571", "71.429", "7.143"),
+    "databricks/databricks-gpt-5-5-pro": ("428.571", "2571.429", "428.571", "428.571"),
     "databricks/databricks-gpt-5-4": ("35.714", "214.286", "35.714", "3.571"),
     "databricks/databricks-gpt-5-4-mini": ("10.714", "64.286", "10.714", "1.071"),
     "databricks/databricks-gpt-5-4-nano": ("2.857", "17.857", "2.857", "0.286"),
+    "databricks/databricks-gemini-3-6-flash": ("26.786", "133.929", "26.786", "2.679"),
+    "databricks/databricks-gemini-3-5-flash": ("26.786", "160.714", "26.786", "2.679"),
+    "databricks/databricks-gemini-3-5-flash-lite": ("5.357", "44.643", "5.357", "0.536"),
     "databricks/databricks-gemini-3-1-pro": ("35.714", "214.286", "35.714", "3.571"),
     "databricks/databricks-gemini-3-pro": ("35.714", "214.286", "35.714", "3.571"),
     "databricks/databricks-gemini-3-flash": ("8.929", "53.571", "8.929", "0.893"),
@@ -65,6 +78,13 @@
     "databricks/databricks-deepseek-v4-flash-0731": ("2.000", "4.000", "2.000", "0.400"),
     "databricks/databricks-deepseek-v4-pro-0813": ("18.857", "56.571", "18.857", "1.886"),
     "databricks/databricks-glm-5-2": ("20.000", "62.857", "20.000", "3.714"),
+    "databricks/databricks-glm-5-3": ("20.000", "62.857", "20.000", "3.714"),
+    "databricks/databricks-glm-5-3-flash": ("2.143", "7.143", "2.143", "0.429"),
+    "databricks/databricks-inkling": ("14.286", "57.857", "14.286", "2.429"),
+    "databricks/databricks-grok-4-6": ("35.714", "107.143", "35.714", "8.929"),
+    "databricks/databricks-qwen35-122b-a10b": ("3.143", "31.429", "3.143", "3.143"),
+    "databricks/databricks-qwen3-next-80b-a3b-instruct": ("2.143", "17.143", "2.143", "2.143"),
+    "databricks/databricks-qwen3-embedding-0-6b": ("0.286", "0", "0.286", "0.286"),
 }
 PROMOTIONAL_DISCOUNT: Final = 0.80
 PROMOTION_EXPIRES: Final = "2027-01-31"
@@ -73,6 +93,10 @@
     "databricks/databricks-gemini-2-5-flash",
 )
 ENTRIES_STORING_LIST_RATE_DESPITE_PROMOTION: Final = (
+    "databricks/databricks-gemini-3-6-flash",
+    "databricks/databricks-gemini-3-5-flash",
+    "databricks/databricks-gemini-3-5-flash-lite",
+    "databricks/databricks-grok-4-6",
     "databricks/databricks-gemini-3-1-pro",
     "databricks/databricks-gemini-3-pro",
     "databricks/databricks-gemini-3-flash",

tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py

@@ -15,6 +15,7 @@
 from litellm.llms.base_llm.ocr.transformation import OCRPage, OCRResponse, OCRUsageInfo
 
 OCR4_COST_PER_PAGE = 0.004
+OCR4_ANNOTATION_COST_PER_PAGE = 0.005
 
 REPO_ROOT = Path(__file__).parents[5]
 MAIN_COST_MAP = REPO_ROOT / "model_prices_and_context_window.json"
@@ -133,3 +134,16 @@ def test_azure_doc_ai_annotation_pages_fall_back_to_ocr_rate(local_model_cost_ma
         call_type="ocr",
     )
     assert cost == pytest.approx(AZURE_DOC_AI_COST_PER_PAGE)
+
+
+def test_azure_ocr4_bills_ocr_and_annotation_pages_at_their_own_rates(local_model_cost_map) -> None:
+    info = litellm.get_model_info(model="azure_ai/mistral-ocr-4-0", custom_llm_provider="azure_ai")
+    assert info["ocr_cost_per_page"] == OCR4_COST_PER_PAGE
+    assert info["annotation_cost_per_page"] == OCR4_ANNOTATION_COST_PER_PAGE
+    cost = completion_cost(
+        completion_response=_annotated_ocr_response("mistral-ocr-4-0", 2, 3),
+        model="azure_ai/mistral-ocr-4-0",
+        custom_llm_provider="azure_ai",
+        call_type="ocr",
+    )
+    assert cost == pytest.approx(2 * OCR4_COST_PER_PAGE + 3 * OCR4_ANNOTATION_COST_PER_PAGE)

tests/test_litellm/proxy/spend_tracking/test_savings.py

@@ -852,7 +852,7 @@ def test_the_served_arm_is_read_from_the_record_not_repriced():
 @pytest.mark.parametrize(
     "basis, expected_multiplier",
     [
-        pytest.param({"service_tier": "priority"}, 2.0, id="priority tier doubles the baseline"),
+        pytest.param({"service_tier": "priority"}, 2.5, id="priority tier uplifts the baseline"),
         pytest.param({"data_residency": "eu"}, 1.1, id="eu residency uplifts the baseline"),
         pytest.param({}, 1.0, id="no basis recorded prices at standard"),
         pytest.param(None, 1.0, id="row predating the field prices at standard"),
@@ -872,7 +872,8 @@ def test_the_baseline_is_priced_on_the_basis_the_request_was_billed_at(basis, ex
     """
     gpt = litellm.get_model_info("gpt-5.5", "openai")
     haiku = litellm.get_model_info("claude-haiku-4-5", "anthropic")
-    assert gpt.get("input_cost_per_token_priority") == 2 * gpt["input_cost_per_token"]
+    assert gpt.get("input_cost_per_token_priority") == pytest.approx(2.5 * gpt["input_cost_per_token"])
+    assert gpt.get("output_cost_per_token_priority") == pytest.approx(2.5 * gpt["output_cost_per_token"])
     assert gpt.get("regional_processing_uplift_multiplier_eu") == 1.1
     assert haiku.get("input_cost_per_token_priority") is None, "served model must not move with the basis"
     assert haiku.get("regional_processing_uplift_multiplier_eu") is None