feat: enable support for non-tool structured outputs on Anthropic Claude Opus 4.5 and 4.6 (use output_format param)

[kelvin-tran]

Relevant issues

Fixes #20533

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

Add Anthropic Opus 4.5 and 4.6 to native structured outputs support

Summary

Add opus-4.5 and opus-4.6 model variants to the set of Anthropic models that use native structured outputs (output_format) instead of the tool-based workaround
Add unit tests verifying Opus 4.5/4.6 use output_format and older models still use the tool-based fallback

Test plan

• test_opus_uses_native_structured_output — parametrized across 4 model name variants, verifies output_format is set and no tool workaround is created
• test_non_structured_output_model_uses_tool_workaround — verifies older models still use tool+tool_choice approach
• Existing structured output tests continue to pass

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 6, 2026 5:25pm

[greptile-apps]

Greptile Overview

Greptile Summary

Added native structured output support for Anthropic Claude Opus 4.5 and 4.6 models by including their model name variants in the output_format parameter check.

• Extended the model substring check to include opus-4.5, opus-4-5, opus-4.6, and opus-4-6
• Models matching these patterns now use Anthropic's native output_format parameter instead of the tool-based workaround
• Added comprehensive test coverage with parametrized tests for all 4 model name variants
• Included regression test to ensure older models still use the tool-based fallback
• poetry.lock changes are purely formatting updates from Poetry version change (2.3.2 → 2.2.1)

Confidence Score: 5/5

• This PR is safe to merge with no risks identified
• The change is minimal and surgical - only adds 4 new string patterns to an existing conditional check. The implementation follows the exact same pattern as existing opus-4.1 support. Comprehensive test coverage validates both the new functionality and ensures backward compatibility with older models.
• No files require special attention

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/transformation.py	Added opus-4.5 and opus-4.6 variants to native structured output support
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py	Added comprehensive tests for Opus 4.5/4.6 native structured outputs and fallback behavior

Sequence Diagram

sequenceDiagram
    participant User
    participant LiteLLM
    participant AnthropicConfig
    participant Anthropic API

    User->>LiteLLM: Request with response_format
    LiteLLM->>AnthropicConfig: map_openai_params(model, response_format)
    
    alt Opus 4.5/4.6 or Sonnet 4.5
        AnthropicConfig->>AnthropicConfig: Check if model contains opus-4.5/opus-4-5/opus-4.6/opus-4-6
        AnthropicConfig->>AnthropicConfig: map_response_format_to_anthropic_output_format()
        AnthropicConfig-->>LiteLLM: Return params with output_format (native)
        LiteLLM->>Anthropic API: Call with output_format
    else Older models (e.g., Sonnet 3.5)
        AnthropicConfig->>AnthropicConfig: map_response_format_to_anthropic_tool()
        AnthropicConfig-->>LiteLLM: Return params with tools + tool_choice (workaround)
        LiteLLM->>Anthropic API: Call with tool-based structured output
    end
    
    Anthropic API-->>LiteLLM: Structured response
    LiteLLM-->>User: Formatted output

Loading

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[kelvin-tran]

Initially closed in favor of #20518, but realize that PR may take longer to push through given complexity. Also this issue was missed in @Sameerlite's #20551 (which aimed to fully support Opus 4.6)

[greptile-apps]

Greptile Overview

Greptile Summary

• Extends Anthropic response_format handling to treat Opus 4.5/4.6 model variants as supporting native structured outputs via output_format, instead of the existing tool+tool_choice workaround.
• Adds unit tests asserting Opus 4.5/4.6 map response_format → output_format (and do not emit tools), while older Claude models still use the tool workaround.
• Includes a large, unrelated poetry.lock regeneration/formatting change that should not be part of this feature PR.

Confidence Score: 3/5

• Feature logic looks safe, but the lockfile changes make the PR risky to merge as-is.
• The functional code change is small and covered by targeted tests, but poetry.lock has been broadly regenerated (different Poetry version header, widespread marker/constraint drift) which can break dependency resolution and is not required for the feature; this should be reverted or regenerated consistently before merge.
• poetry.lock

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/transformation.py	Adds Opus 4.5/4.6 model substrings to the allowlist for mapping OpenAI response_format to Anthropic native structured outputs via output_format; logic change is localized to response_format handling.
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py	Adds unit tests asserting Opus 4.5/4.6 use output_format (no tool workaround) and that older models still use tools/tool_choice; removes an unused TestClient import.
poetry.lock	Lockfile regenerated with a different Poetry version and includes a resolved merge conflict section; large unrelated marker/constraint changes make this file unsafe as-is and likely to break dependency resolution.

Sequence Diagram

sequenceDiagram
    participant Caller as LiteLLM caller
    participant AC as AnthropicConfig.map_openai_params
    participant RF as response_format mapping

    Caller->>AC: non_default_params{response_format}, model
    AC->>RF: if model contains opus-4.5/4.6 or sonnet/opus allowlist
    alt Allowlisted model (Opus 4.5/4.6)
        RF-->>AC: map_response_format_to_anthropic_output_format()
        AC-->>Caller: optional_params.output_format set
    else Other models
        RF-->>AC: map_response_format_to_anthropic_tool()
        AC-->>Caller: optional_params.tools + optional_params.tool_choice set
    end
    AC-->>Caller: optional_params.json_mode = true

Loading

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}