fix(bedrock): preserve output_config.effort for Claude 4.6 on Invoke API

[KevinZhao]

Summary

Bedrock Invoke API previously stripped output_config unconditionally for all models (#22797). However, Claude 4.6 models (Opus 4.6 and Sonnet 4.6) support output_config.effort as a stable API parameter for controlling adaptive thinking effort level.

This PR makes the stripping conditional:

• Claude 4.6+ models: preserve output_config so effort (low/medium/high/max) is forwarded to Bedrock
• Older models: continue stripping output_config to avoid request rejection

Changes

• litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py — conditional output_config pop for /v1/chat/completions → Bedrock Invoke path
• litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py — same for /v1/messages → Bedrock Invoke path
• litellm/types/llms/anthropic.py — add "max" to AnthropicOutputConfig.effort Literal type (Opus 4.6 exclusive)
• Added tests for both preservation (4.6) and stripping (pre-4.6) scenarios on both paths

Test plan

• test_bedrock_chat_invoke_preserves_output_config_for_claude_4_6 — Sonnet 4.6 effort=high, Opus 4.6 effort=max
• test_bedrock_chat_invoke_strips_output_config_for_pre_4_6 — Sonnet 4.5, Opus 4.5
• test_bedrock_messages_preserves_output_config_for_claude_4_6 — Opus 4.6 effort=max, Sonnet 4.6 effort=high
• test_bedrock_messages_strips_output_config_for_pre_4_6 — Claude 3 Haiku, Sonnet 4, Opus 4.5
• Existing test_output_config_removed_from_bedrock_chat_invoke_request still passes (regression)

pytest tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py -v
pytest tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py -v

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 18, 2026 7:33am

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR makes the Bedrock Invoke API's output_config stripping conditional: Claude 4.6+ models (Opus 4.6, Sonnet 4.6) preserve output_config.effort so it is forwarded to Bedrock, while pre-4.6 models continue to have it stripped to avoid request rejection. It also expands the AnthropicOutputConfig.effort Literal type to include "max" (Opus 4.6 exclusive).

Key changes:

• litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py — output_config pop is now conditional on AnthropicConfig._is_claude_4_6_model(model)
• litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py — same conditional guard via AnthropicModelInfo._is_claude_4_6_model(model)
• litellm/types/llms/anthropic.py — adds "max" to the effort Literal
• Tests added for both paths (preserve for 4.6, strip for pre-4.6)

Concern:

• Both transformation files use _is_claude_4_6_model, a hardcoded model-name string check. Per AGENTS.md §8 and the project's custom instruction, model capability flags must live in model_prices_and_context_window.json and be read via get_model_info (or an existing helper like supports_reasoning). Claude 4.6 models already carry "supports_reasoning": true in that JSON, so a data-driven guard is achievable today without adding new fields. Future models that gain output_config.effort support on Bedrock will require a code change here before users benefit, unless this is refactored to be config-driven.

Confidence Score: 2/5

• The fix is functionally correct but relies on a hardcoded model-name check that violates the project's own architectural rule — future models won't benefit automatically.
• The core logic (conditional strip vs preserve) is sound and the tests are thorough mock-only unit tests. However, _is_claude_4_6_model is a hardcoded string-matching function, which directly contradicts AGENTS.md §8 and custom rule 2605a1b1. The rule exists to prevent exactly this: users having to wait for a LiteLLM upgrade when a future model gains output_config.effort support on Bedrock Invoke. A data-driven approach using supports_reasoning (already true for all 4.6 models in the JSON) or a new supports_output_config_effort field would be compliant and more future-proof.
• Both transformation files need attention: litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py (line 113) and litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py (line 425).

Important Files Changed

Filename	Overview
litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py	Conditionally preserves output_config for Claude 4.6 models using _is_claude_4_6_model — a hardcoded string check that violates the project rule against model-specific flags outside model_prices_and_context_window.json.
litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py	Same hardcoded _is_claude_4_6_model violation as the chat path; additionally the comment numbering (5b appears before 5a) is inverted.
litellm/types/llms/anthropic.py	Adds "max" to the AnthropicOutputConfig.effort Literal — a straightforward type-system fix that is accurate and non-breaking.
tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py	Adds well-structured unit tests for both preservation and stripping paths; all calls are mocked (no real network calls).
tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py	Adds equivalent coverage for the messages Invoke path; tests are mock-only and complete.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["litellm /v1/chat/completions or /v1/messages\n(with output_config.effort)"] --> B{Bedrock Invoke\ntransformation}

    B --> C{_is_claude_4_6_model?}

    C -- "Yes (Opus/Sonnet 4.6)" --> D["Keep output_config\nin request body"]
    C -- "No (pre-4.6)" --> E["Pop output_config\nfrom request body"]

    D --> F["Bedrock Invoke API\n→ effort forwarded"]
    E --> G["Bedrock Invoke API\n→ effort stripped"]

    style D fill:#c8f7c5,stroke:#27ae60
    style E fill:#fdecea,stroke:#e74c3c
    style C fill:#fef9e7,stroke:#f39c12

Loading

_{Last reviewed commit: "fix(bedrock): preser..."}

[greptile-apps]

Hardcoded model check violates AGENTS.md rule #8

_is_claude_4_6_model is a hardcoded string-match check. Per the project's custom rule and AGENTS.md §8 ("Do not hardcode model-specific flags"), model capability decisions must be driven by model_prices_and_context_window.json read via get_model_info (or an existing helper like supports_reasoning). Without this, any future model that gains output_config.effort support on Bedrock Invoke requires a code change here before users benefit — exactly the problem the rule was written to avoid.

All Claude 4.6 models already carry "supports_reasoning": true in the model-prices JSON, so supports_reasoning (or a new supports_output_config_effort flag added to the JSON) could drive this decision without touching the code for future models.

# Option A — reuse existing supports_reasoning helper (reads from model_prices_and_context_window.json)
from litellm.utils import supports_reasoning

if not (AnthropicConfig._is_claude_4_6_model(model) or supports_reasoning(model=model, custom_llm_provider=self.custom_llm_provider)):
    _anthropic_request.pop("output_config", None)

This same violation also exists at litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py:425.

Rule Used: What: Do not hardcode model-specific flags in the ... (source)

[greptile-apps]

Comment numbering out of order (5b before 5a)

The comment label # 5b. on line 422 appears in the file before # 5a. on line 428, which reverses the expected ordering and makes the numbering misleading when reading the code sequentially.

Suggested change

	# 5a. Remove `custom` field from tools (Bedrock doesn't support it)
	# 5a. Remove `custom` field from tools (Bedrock doesn't support it)

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing KevinZhao:fix/bedrock-invoke-output-config-effort-4-6 (34587d5) with main (cec3e9e)}

[KevinZhao]

CLA signed