Merge pull request #38615 from BerriAI/litellm_mantle_gpt5_cost_tiers

fix(model_prices): add bedrock_mantle gpt-5.5/5.4 272K tiers, align sol with AWS invoice
2026-09-06 08:16:43 +00:00 · 2026-08-28 10:09:52 -07:00 · 2026-08-28 10:09:52 -07:00 · 3a52ae0a3f
commit 3a52ae0a3f
parent e1cc96eaa7 837bcba32d
4 changed files with 82 additions and 39 deletions
--- a/litellm/model_prices_and_context_window_backup.json
+++ b/litellm/model_prices_and_context_window_backup.json
@ -49829,14 +49829,14 @@
        "supports_tool_choice": true
    },
    "bedrock_mantle/openai.gpt-5.6-sol": {
-        "input_cost_per_token": 4.4e-06,
-        "input_cost_per_token_above_272k_tokens": 8.8e-06,
-        "cache_creation_input_token_cost": 5.5e-06,
-        "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05,
-        "cache_read_input_token_cost": 4.4e-07,
-        "cache_read_input_token_cost_above_272k_tokens": 8.8e-07,
-        "output_cost_per_token": 2.2e-05,
-        "output_cost_per_token_above_272k_tokens": 3.3e-05,
+        "input_cost_per_token": 5.5e-06,
+        "input_cost_per_token_above_272k_tokens": 1.1e-05,
+        "cache_creation_input_token_cost": 6.875e-06,
+        "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
+        "cache_read_input_token_cost": 5.5e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
+        "output_cost_per_token": 3.3e-05,
+        "output_cost_per_token_above_272k_tokens": 4.95e-05,
        "litellm_provider": "bedrock_mantle",
        "max_input_tokens": 1050000,
        "max_output_tokens": 128000,
@ -50113,8 +50113,11 @@
    },
    "bedrock_mantle/openai.gpt-5.5": {
        "input_cost_per_token": 5.5e-06,
+        "input_cost_per_token_above_272k_tokens": 1.1e-05,
        "cache_read_input_token_cost": 5.5e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
        "output_cost_per_token": 3.3e-05,
+        "output_cost_per_token_above_272k_tokens": 4.95e-05,
        "litellm_provider": "bedrock_mantle",
        "max_input_tokens": 1050000,
        "max_output_tokens": 128000,
@ -50140,8 +50143,11 @@
    },
    "bedrock_mantle/openai.gpt-5.4": {
        "input_cost_per_token": 2.75e-06,
+        "input_cost_per_token_above_272k_tokens": 5.5e-06,
        "cache_read_input_token_cost": 2.75e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 5.5e-07,
        "output_cost_per_token": 1.65e-05,
+        "output_cost_per_token_above_272k_tokens": 2.475e-05,
        "litellm_provider": "bedrock_mantle",
        "max_input_tokens": 1050000,
        "max_output_tokens": 128000,
--- a/model_prices_and_context_window.json
+++ b/model_prices_and_context_window.json
@ -49829,14 +49829,14 @@
        "supports_tool_choice": true
    },
    "bedrock_mantle/openai.gpt-5.6-sol": {
-        "input_cost_per_token": 4.4e-06,
-        "input_cost_per_token_above_272k_tokens": 8.8e-06,
-        "cache_creation_input_token_cost": 5.5e-06,
-        "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05,
-        "cache_read_input_token_cost": 4.4e-07,
-        "cache_read_input_token_cost_above_272k_tokens": 8.8e-07,
-        "output_cost_per_token": 2.2e-05,
-        "output_cost_per_token_above_272k_tokens": 3.3e-05,
+        "input_cost_per_token": 5.5e-06,
+        "input_cost_per_token_above_272k_tokens": 1.1e-05,
+        "cache_creation_input_token_cost": 6.875e-06,
+        "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
+        "cache_read_input_token_cost": 5.5e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
+        "output_cost_per_token": 3.3e-05,
+        "output_cost_per_token_above_272k_tokens": 4.95e-05,
        "litellm_provider": "bedrock_mantle",
        "max_input_tokens": 1050000,
        "max_output_tokens": 128000,
@ -50113,8 +50113,11 @@
    },
    "bedrock_mantle/openai.gpt-5.5": {
        "input_cost_per_token": 5.5e-06,
+        "input_cost_per_token_above_272k_tokens": 1.1e-05,
        "cache_read_input_token_cost": 5.5e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
        "output_cost_per_token": 3.3e-05,
+        "output_cost_per_token_above_272k_tokens": 4.95e-05,
        "litellm_provider": "bedrock_mantle",
        "max_input_tokens": 1050000,
        "max_output_tokens": 128000,
@ -50140,8 +50143,11 @@
    },
    "bedrock_mantle/openai.gpt-5.4": {
        "input_cost_per_token": 2.75e-06,
+        "input_cost_per_token_above_272k_tokens": 5.5e-06,
        "cache_read_input_token_cost": 2.75e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 5.5e-07,
        "output_cost_per_token": 1.65e-05,
+        "output_cost_per_token_above_272k_tokens": 2.475e-05,
        "litellm_provider": "bedrock_mantle",
        "max_input_tokens": 1050000,
        "max_output_tokens": 128000,
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@ -532,40 +532,71 @@ def test_generic_cost_per_token_bedrock_mantle_gpt56_long_context(_local_model_c


@pytest.mark.parametrize(
-    "model",
+    "model,input_rate,cache_read_rate,output_rate,long_input_rate,long_cache_read_rate,long_output_rate",
    [
-        "bedrock_mantle/openai.gpt-5.5",
-        "bedrock_mantle/openai.gpt-5.4",
+        ("bedrock_mantle/openai.gpt-5.5", 5.5e-06, 5.5e-07, 3.3e-05, 1.1e-05, 1.1e-06, 4.95e-05),
+        ("bedrock_mantle/openai.gpt-5.4", 2.75e-06, 2.75e-07, 1.65e-05, 5.5e-06, 5.5e-07, 2.475e-05),
+        ("bedrock_mantle/openai.gpt-5.6-sol", 5.5e-06, 5.5e-07, 3.3e-05, 1.1e-05, 1.1e-06, 4.95e-05),
    ],
 )
-def test_generic_cost_per_token_bedrock_mantle_gpt55_gpt54_long_context_flat_rate(_local_model_cost_map, model):
-    """Bedrock serves gpt-5.5 and gpt-5.4 up to its enforced 1,050,000-token prompt maximum and documents
-    no long-context tier for them, so a prompt past 272K is billed at the flat per-token rates."""
+def test_generic_cost_per_token_bedrock_mantle_gpt5_matches_aws_invoiced_rates(
+    _local_model_cost_map,
+    model,
+    input_rate,
+    cache_read_rate,
+    output_rate,
+    long_input_rate,
+    long_cache_read_rate,
+    long_output_rate,
+):
+    """AWS bills a Bedrock GPT-5.x prompt past 272K under its long-context usage types, the whole prompt at
+    2x input, 2x cache read, and 1.5x output. The flat rates undercounted a 300K gpt-5.5 prompt by half and
+    sol's base rates sat 20% under the invoice."""

-    model_cost_map = litellm.model_cost[model]
-    assert model_cost_map["max_input_tokens"] == 1050000
-    assert [key for key in model_cost_map if "above_272k" in key] == []
-
-    served_prompt_tokens = 1030590
    cached_tokens = 100000
    completion_tokens = 1000
-    usage = Usage(
-        prompt_tokens=served_prompt_tokens,
+
+    invoiced_prompt_tokens = 300238
+    long_usage = Usage(
+        prompt_tokens=invoiced_prompt_tokens,
        completion_tokens=completion_tokens,
-        total_tokens=served_prompt_tokens + completion_tokens,
+        total_tokens=invoiced_prompt_tokens + completion_tokens,
        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=cached_tokens),
    )
-    prompt_cost, completion_cost = generic_cost_per_token(
+    long_prompt_cost, long_completion_cost = generic_cost_per_token(
        model=model,
-        usage=usage,
+        usage=long_usage,
        custom_llm_provider="bedrock_mantle",
    )
-    assert round(prompt_cost, 10) == round(
-        model_cost_map["input_cost_per_token"] * (served_prompt_tokens - cached_tokens)
-        + model_cost_map["cache_read_input_token_cost"] * cached_tokens,
-        10,
+    assert long_prompt_cost == pytest.approx(
+        long_input_rate * (invoiced_prompt_tokens - cached_tokens) + long_cache_read_rate * cached_tokens
    )
-    assert round(completion_cost, 10) == round(model_cost_map["output_cost_per_token"] * completion_tokens, 10)
+    assert long_completion_cost == pytest.approx(long_output_rate * completion_tokens)
+
+    threshold_prompt_tokens = 272000
+    short_usage = Usage(
+        prompt_tokens=threshold_prompt_tokens,
+        completion_tokens=completion_tokens,
+        total_tokens=threshold_prompt_tokens + completion_tokens,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=cached_tokens),
+    )
+    short_prompt_cost, short_completion_cost = generic_cost_per_token(
+        model=model,
+        usage=short_usage,
+        custom_llm_provider="bedrock_mantle",
+    )
+    assert short_prompt_cost == pytest.approx(
+        input_rate * (threshold_prompt_tokens - cached_tokens) + cache_read_rate * cached_tokens
+    )
+    assert short_completion_cost == pytest.approx(output_rate * completion_tokens)
+
+
+def test_bedrock_mantle_gpt56_sol_cache_write_matches_aws_invoiced_rate(_local_model_cost_map):
+    """The invoice bills sol 30-minute cache writes at $6.88 per million tokens, 1.25x the $5.50 input rate."""
+
+    sol = litellm.model_cost["bedrock_mantle/openai.gpt-5.6-sol"]
+    assert sol["cache_creation_input_token_cost"] == pytest.approx(6.875e-06)
+    assert sol["cache_creation_input_token_cost_above_272k_tokens"] == pytest.approx(1.375e-05)


 def test_generic_cost_per_token_honors_non_standard_above_threshold():
--- a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py
+++ b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py
@ -1695,7 +1695,7 @@ class TestBedrockMantleResponsesPricing:
    @pytest.mark.parametrize(
        "model, input_cost, cache_creation_cost, cache_read_cost, output_cost",
        [
-            ("openai.gpt-5.6-sol", 4.4e-06, 5.5e-06, 4.4e-07, 2.2e-05),
+            ("openai.gpt-5.6-sol", 5.5e-06, 6.875e-06, 5.5e-07, 3.3e-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),
        ],
@ -1718,7 +1718,7 @@ class TestBedrockMantleResponsesPricing:
    @pytest.mark.parametrize(
        "model, input_cost, output_cost",
        [
-            ("openai.gpt-5.6-sol", 4.4e-06, 2.2e-05),
+            ("openai.gpt-5.6-sol", 5.5e-06, 3.3e-05),
            ("openai.gpt-5.6-terra", 2.2e-06, 1.32e-05),
            ("openai.gpt-5.6-luna", 2.2e-07, 1.32e-06),
        ],