fix(anthropic): drop unsupported speed param with drop_params

[krrish-berri-2]

Relevant issues

Fixes requests to Anthropic models that do not support fast mode (e.g. claude-sonnet-4-6) failing with a 400 when callers pass speed and have drop_params enabled

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

With drop_params=True, /v1/messages to a non-Opus model no longer forwards speed to Anthropic.

# start proxy (separate terminal)
python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver

# before fix: Anthropic 400 "speed: Extra inputs are not permitted"
# after fix: request succeeds; speed is dropped with a verbose warning
curl -s http://localhost:4000/v1/messages \
  -H "Authorization: Bearer sk-1234" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4-6",
    "max_tokens": 16,
    "messages": [{"role": "user", "content": "hi"}],
    "speed": "fast"
  }' | jq .

Unit tests:

LITELLM_LOCAL_MODEL_COST_MAP=True python -m pytest \
  tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py -k "speed or fast_mode" \
  tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_speed.py \
  tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_request_optional_param_utils.py -q

Type

🐛 Bug Fix

Changes

Anthropic fast mode (speed) is a research preview on the direct Claude API only, for Opus 4.6, 4.7, and 4.8. It is not available on Bedrock, Vertex, or Azure Foundry.

This PR gates speed on a new supports_speed flag in the model map, scoped to the direct Anthropic provider. The lookup reads the exact routed model-map entry (no provider alias walk). Vertex, Azure, and Bedrock reuse the shared Anthropic transform after stripping their provider prefix, so a bare claude-opus-4-8 arriving from one of those routes would otherwise resolve to the fast-mode-capable direct entry and forward speed upstream; the gate now also checks custom_llm_provider so those routes are kept out. When the model does not support speed and litellm.drop_params or per-request drop_params is true, speed is stripped before the upstream request on both the chat completions path and the /v1/messages passthrough. When drop_params is false, an UnsupportedParamsError is raised instead.

Model map entries updated for claude-opus-4-6, claude-opus-4-6-20260205, claude-opus-4-7, claude-opus-4-7-20260416, and claude-opus-4-8 in both model_prices_and_context_window.json and the bundled backup. Bedrock/Vertex/Azure Opus entries are intentionally unchanged.

Made with Cursor

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ mateo-berri
❌ krrish-berri-2
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fixes Anthropic API 400 errors when callers pass speed to models that don't support fast mode (e.g. claude-sonnet-4-6) and have drop_params enabled. It gates the speed parameter behind a new supports_speed flag in the model cost map, read via an exact-lookup helper that deliberately skips provider-alias walking so Bedrock/Vertex/Azure Opus entries cannot inherit the flag from the direct-API entry.

• Adds _get_exact_model_capability / _model_supports_speed_param / _maybe_drop_speed_param to AnthropicConfig; the drop-or-raise logic is invoked in both map_openai_params and (as a safety net) transform_request, and is also wired into the /v1/messages passthrough path via get_requested_anthropic_messages_optional_param.
• Sets supports_speed: true on exactly five direct-API Opus entries (claude-opus-4-6, claude-opus-4-6-20260205, claude-opus-4-7, claude-opus-4-7-20260416, claude-opus-4-8) in both the primary and backup JSON; all Bedrock/Vertex/Azure variants are intentionally left unchanged.
• Test coverage spans both paths (chat completions and passthrough), both drop_params states (drop vs. raise), and the Vertex/Azure/Bedrock provider-gating regression.

Confidence Score: 5/5

Safe to merge. The change is well-scoped: it only affects requests that include a speed parameter, and only for models that don't support fast mode. Existing Opus callers with speed on the direct Anthropic API are unaffected; all other providers are correctly blocked at the custom_llm_provider check before any model-map lookup occurs.

The fix correctly reads capability data from the model cost JSON rather than hardcoding model names in Python, both paths (chat completions and /v1/messages passthrough) are handled, and the provider guard prevents the Bedrock/Vertex/Azure prefix-stripping bypass. No existing tests were weakened, and the new tests cover the full decision matrix.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/common_utils.py	Adds _get_exact_model_capability – a deliberate non-alias-walking counterpart to _get_model_capability – used to gate speed on the exact model-map entry only, preventing Bedrock/Vertex prefix-stripped names from accidentally resolving to a fast-mode-capable direct-API entry.
litellm/llms/anthropic/chat/transformation.py	Adds _model_supports_speed_param and _maybe_drop_speed_param; calls the latter inside the speed branch of map_openai_params (emits warning and pops on drop, raises on no-drop) and again in transform_request as a safety net for speed injected outside the normal param-mapping flow.
litellm/llms/anthropic/experimental_pass_through/messages/utils.py	Extends get_requested_anthropic_messages_optional_param with keyword-only model, drop_params, and custom_llm_provider params; delegates to _maybe_drop_speed_param so the passthrough path raises or strips speed consistently with the chat path.
litellm/llms/anthropic/experimental_pass_through/messages/handler.py	Wires model, drop_params, and custom_llm_provider through to get_requested_anthropic_messages_optional_param; change is minimal and correct.
model_prices_and_context_window.json	Adds supports_speed: true exactly to the five direct-API Anthropic Opus entries (claude-opus-4-6, claude-opus-4-6-20260205, claude-opus-4-7, claude-opus-4-7-20260416, claude-opus-4-8). Bedrock, Vertex, Azure, and other provider-specific entries are intentionally left unchanged.
litellm/model_prices_and_context_window_backup.json	Mirrors the same five supports_speed: true additions as the primary JSON; both files stay in sync.
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py	Adds tests for supported/rejected models, provider-gating (vertex_ai/azure_ai/bedrock), drop_params=True strips unsupported speed, drop_params=False raises UnsupportedParamsError; existing test_fast_mode_parameter_mapping is unchanged and continues to verify speed passes through for Opus models.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_speed.py	New test file covering the passthrough /v1/messages path: drops speed for non-Opus, keeps it for Opus, raises when drop_params=False, and verifies vertex_ai Opus is also blocked.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_request_optional_param_utils.py	Extends existing regression tests for the param-filter fast path with two new cases verifying global litellm.drop_params behaviour in the utils layer.
tests/test_litellm/test_utils.py	Adds supports_speed to the JSON schema validator so the new field is recognized as a valid boolean key; no logic changes.

_{Reviews (3): Last reviewed commit: "fix(anthropic): gate speed param by rout..." | Re-trigger Greptile}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f7cc14505f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[krrish-berri-2]

@greptile review

[mateo-berri]

@greptileai review

---

Generated by Claude Code

[mateo-berri]

bugbot run

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit bc36f4a. Configure here.}

[mateo-berri]

LGTM; thanks!