Migrate to Opus 4.6

[12458]

Relevant issues

Fixes #20519

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• 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

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

New Feature

Changes

Migrate Claude Opus 4.6 from the deprecated thinking: {type: "enabled", budget_tokens: N} to thinking: {type: "adaptive"} + output_config: {effort: "high"|"medium"|"low"}. The effort-2025-11-24 beta header is no longer needed for Opus 4.6 since effort is GA.

Summary of changes

• litellm/types/llms/anthropic.py — AnthropicThinkingParam.type now accepts "adaptive" in addition to "enabled"
• litellm/llms/base_llm/chat/transformation.py — is_thinking_enabled recognizes type: "adaptive"
• litellm/llms/anthropic/chat/transformation.py — Core logic:
  • Added _is_claude_opus_4_6() model detection
  • Added _get_adaptive_thinking_param() and _budget_tokens_to_effort() helpers
  • map_openai_params: for Opus 4.6, reasoning_effort maps to adaptive thinking + effort; deprecated thinking: {type: "enabled"} is auto-converted to adaptive
  • Added Opus 4.6 to the response_format native output_format model list
• litellm/llms/anthropic/common_utils.py — is_effort_used returns False for Opus 4.6 (no beta header needed)
• litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py — Added Opus 4.6 to extended thinking support on Bedrock
• litellm/llms/bedrock/chat/converse_transformation.py — Bedrock Converse uses adaptive thinking for Opus 4.6

Test plan

Added 16 tests in tests/test_litellm/llms/anthropic/chat/test_anthropic_opus_4_6_adaptive_thinking.py:

• reasoning_effort → adaptive thinking + effort mapping (high, medium, low, minimal)
• Deprecated thinking: {type: "enabled"} auto-conversion for Opus 4.6
• thinking: {type: "adaptive"} passthrough
• _budget_tokens_to_effort threshold mapping
• Opus 4.5 behavior unchanged
• Beta header not set for Opus 4.6, still set for Opus 4.5
• is_thinking_enabled recognizes adaptive type
• No auto max_tokens adjustment for adaptive (no budget_tokens)
• Opus 4.6 response_format uses native output_format
• Model name detection for various formats
• Bedrock Converse adaptive thinking for Opus 4.6

Linting

https://github.com/BerriAI/litellm/actions/runs/21727820552/job/62674426851

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 5, 2026 8:46pm

[greptile-apps]

Greptile Overview

Greptile Summary

This PR updates Anthropic/Claude parameter mapping to support Opus 4.6’s GA “adaptive thinking” API by:

• Detecting Opus 4.6 model name variants and mapping reasoning_effort to thinking: {type: "adaptive"} plus output_config: {effort: ...}.
• Auto-converting deprecated thinking: {type: "enabled", budget_tokens: N} to adaptive thinking for Opus 4.6 (deriving effort from budget_tokens when present).
• Ensuring is_thinking_enabled treats thinking.type="adaptive" as reasoning-enabled.
• Skipping the effort beta header for Opus 4.6 (effort treated as GA) while keeping Opus 4.5 behavior unchanged.
• Extending Bedrock transformations to recognize Opus 4.6 for extended/adaptive thinking.

New tests exercise the mapping logic, model detection, beta header behavior, and Bedrock Converse handling for Opus 4.6.

Confidence Score: 4/5

• This PR looks safe to merge with low functional risk given the targeted parameter mapping changes and added test coverage.
• Changes are mostly additive and gated on Opus 4.6 model detection; existing Opus 4.5 and non-Opus paths appear preserved. I couldn’t execute the test suite in this environment (no python/pytest available), so confidence is not 5/5.
• tests/test_litellm/llms/anthropic/chat/test_anthropic_opus_4_6_adaptive_thinking.py (test runner assumptions); litellm/llms/anthropic/chat/transformation.py (Opus 4.6 detection + mapping logic)

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/transformation.py	Adds Opus 4.6 model detection and maps reasoning_effort / deprecated thinking(enabled) to adaptive thinking + output_config.effort; includes Opus 4.6 in native output_format allowlist.
litellm/llms/anthropic/common_utils.py	Updates beta-header feature detection; treats Opus 4.6 as not requiring the effort beta header by short-circuiting is_effort_used(model).
litellm/llms/base_llm/chat/transformation.py	Extends is_thinking_enabled to recognize thinking.type='adaptive' in addition to 'enabled'.
litellm/llms/bedrock/chat/converse_transformation.py	Bedrock Converse reasoning_effort mapping uses adaptive thinking for Opus 4.6 model strings; otherwise keeps prior enabled-thinking mapping.
litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py	Adds Opus 4.6 model patterns to Bedrock extended thinking supported_patterns list.
litellm/types/llms/anthropic.py	Extends AnthropicThinkingParam.type literal to include 'adaptive' and minor formatting tweaks/newline fix.
poetry.lock	Updates locked dependency versions (e.g., aiohttp 3.13.2→3.13.3).
tests/test_litellm/llms/anthropic/chat/test_anthropic_opus_4_6_adaptive_thinking.py	Adds unit tests for Opus 4.6 adaptive thinking/effort mapping and beta header behavior; modifies sys.path at runtime which may break test runner assumptions.

Sequence Diagram

sequenceDiagram
    participant Client
    participant LiteLLM as LiteLLM (OpenAI params)
    participant ATrans as AnthropicConfig.map_openai_params
    participant Base as BaseConfig.is_thinking_enabled
    participant Headers as AnthropicModelInfo.get_anthropic_headers
    participant Bedrock as AmazonConverseConfig._handle_reasoning_effort_parameter
    participant Anthropic as Anthropic API

    Client->>LiteLLM: ChatCompletion(model=opus-4-6, reasoning_effort/thinking)

    LiteLLM->>ATrans: map_openai_params(non_default_params, optional_params, model)
    ATrans->>Base: is_thinking_enabled(non_default_params)
    alt model is Opus 4.6 && reasoning_effort provided
        ATrans-->>ATrans: set thinking={type:"adaptive"}
        ATrans-->>ATrans: set output_config={effort: low|medium|high}
    else model is Opus 4.6 && thinking.type=="enabled"
        ATrans-->>ATrans: convert thinking to {type:"adaptive"}
        ATrans-->>ATrans: optionally derive effort from budget_tokens
    else other models
        ATrans-->>ATrans: keep existing thinking mapping
    end

    LiteLLM->>Headers: get_anthropic_headers(optional_params, model)
    alt model is Opus 4.6
        Headers-->>Headers: is_effort_used() returns false (no effort beta header)
    else effort used on other models
        Headers-->>Headers: add anthropic-beta: effort-2025-11-24
    end

    LiteLLM->>Anthropic: POST /messages (thinking + output_config)
    Anthropic-->>LiteLLM: Response

    %% Bedrock path
    Client->>LiteLLM: Bedrock Converse call (model contains opus-4-6)
    LiteLLM->>Bedrock: _handle_reasoning_effort_parameter(model, reasoning_effort)
    alt Opus 4.6
        Bedrock-->>Bedrock: set thinking={type:"adaptive"}
    else
        Bedrock-->>Bedrock: set thinking via _map_reasoning_effort
    end

Loading

[greptile-apps]

_{8 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[krrishdholakia]

@Sameerlite do your PR's already address this?

[Sameerlite]

Covered all this here: #20551
Closing this