Skip to content

fix(openai): honour a pinned temperature, warn when dropped; add gpt-6.1-sol - #1000

Merged
sroussey merged 3 commits into
mainfrom
claude/intelligent-ride-qcx8o9
Sep 30, 2026
Merged

sroussey merged 3 commits into
mainfrom
claude/intelligent-ride-qcx8o9

Conversation

@sroussey

@sroussey sroussey commented Sep 30, 2026 •

Copy link
Copy Markdown
Collaborator

Problem

finalizeResponsesRequest deleted params.temperature whenever reasoning.effort !== "none". Since unconfigured reasoning models default to effort medium, a caller pinning temperature: 0 had it silently ignored.

Change

  • Pinned temperature, no effort configured, GPT-5.6 model (the family verified to accept effort none; class policy and effort_options must allow it): send reasoning.effort: "none" and keep the temperature.
  • Every other reasoning model (gpt-6, gpt-6.1, gpt-5.5 and earlier, o-series): the request is not guessed at. The default effort is sent, the temperature is dropped, and a warning is logged once per model and effort (warnTemperatureDroppedOnce).
  • gpt-6 never receives none, however the record asks for it.
  • Adds gpt-6.1-sol to the fallback model list and pricing (short $2 / cached $0.10 / cache-write $2.50 / output $10; long-context tier derived as 2x/1.5x, which matches the published long row). It resolves to the gpt-6 effort policy (low..max, no none).
  • Other OpenAI pricing was compared with the published page and already matched. Nothing else changed.

Tests

Tests are added for gpt-5.6 (kept), gpt-6 and gpt-6.1-sol, gpt-5.5, o3, gpt-4o, and for gpt-6.1-sol pricing and effort policy. 153 tests pass across the 9 affected files. bun run format and bun run format-check are clean and touched no unrelated files.

Not verified

OpenAI's docs page did not state which models reject temperature or top_p, so that behaviour is unchanged for every model except the gpt-5.6 path above.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TzDPzEBWQ9BxV7Pwfo5Db7

An unconfigured reasoning model now defaults to effort medium, which made
Responses drop a pinned temperature silently. Where the model allows effort
none and none is configured, keep reasoning off so the temperature is sent;
otherwise warn once per model when it is dropped.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TzDPzEBWQ9BxV7Pwfo5Db7
…ath to GPT-5.6

Prices for gpt-6.1-sol come from OpenAI's published pricing page. A pinned
temperature keeps reasoning off only on gpt-5.6 models; every other reasoning
model drops it with a warning instead of sending an effort that may 400.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TzDPzEBWQ9BxV7Pwfo5Db7
@sroussey sroussey changed the title fix(openai): honour a pinned temperature and warn when it is dropped fix(openai): honour a pinned temperature, warn when dropped; add gpt-6.1-sol Sep 30, 2026

Copy link
Copy Markdown
Collaborator Author

The build check failed at bun install, before any build or test step: the onnxruntime-node postinstall got ETIMEDOUT fetching its binary. That's a runner network failure and doesn't touch this diff. I tried to re-run the failed job but GitHub refused because the run is still in progress, so I'll re-run it once the run completes.


Generated by Claude Code

…warning

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TzDPzEBWQ9BxV7Pwfo5Db7
@sroussey
sroussey merged commit cfe01fe into main Sep 30, 2026
18 checks passed
@sroussey
sroussey deleted the claude/intelligent-ride-qcx8o9 branch September 30, 2026 15:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants