Skip to content

🤖 feat: expose GPT-5.6 pro reasoning mode as a thinking-slider toggle #3704

Description

@ammar-agent

Summary

GPT-5.6 (Sol/Terra/Luna) supports reasoning.mode: "pro" on the Responses API: the model performs more work to improve reliability on difficult tasks and returns a single final answer. OpenAI recommends it when quality matters more than latency and token usage. Mux does not expose it yet — this is a known gap from #3702, which shipped GPT-5.6 support with the native max reasoning effort only.

Proposed UX

A toggle inside the thinking slider, shown only when the selected model supports pro mode (GPT-5.6 family). Pro mode is orthogonal to effort (effort: high, mode: pro is valid), so it must not be modeled as a new ThinkingLevel — it composes with whatever level is selected.

Implementation notes (from the reverted spike in #3702)

  • The @ai-sdk/openai Responses schema only builds reasoning.effort/reasoning.summary; it cannot send reasoning.mode. A working approach mirrored the Anthropic xhigh effort rewrite: emit a Mux-internal header in buildRequestHeaders and have the OpenAI fetch wrapper strip it + inject reasoning.mode into the body on the wire (see 🤖 feat: add GPT-5.6 Sol/Terra/Luna models and move gpt alias to Sol #3702 history, commit 991009e, later reverted).
  • Gates needed: GPT-5.6 family only (openaiSupportsNativeMaxEffort covers the same family), direct OpenAI routes only (the OpenAI fetch wrapper is not applied on passthrough gateways), Responses wire format only.
  • Codex review caveat: headless createModel() callers (title generation, advisor streams) never call buildRequestHeaders, so a provider-level config knob needs a fallback path in the wrapper (or explicit plumbing) to apply there.
  • Live-verified July 9, 2026: reasoning: {"effort": "high", "mode": "pro"} completes on gpt-5.6-sol with a plain API key.
  • Worth checking whether the AI SDK adds native reasoning.mode support before building the wire-level workaround.

Generated with mux • Model: anthropic:claude-fable-5 • Thinking: high • Cost: $18.71

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions