`thinking="minimal"` sends unsupported reasoning effort to GPT-5.6 models

[daikeren]

Initial Checks

• I'm using the latest version of Pydantic AI
• I've searched for my issue in the issue tracker before opening this issue

Description

The unified thinking='minimal' setting maps directly to OpenAI's reasoning_effort='minimal'. All three GPT-5.6 variants reject that value with HTTP 400; their lowest supported active effort is low.

Live requests against both the Responses and Chat Completions APIs confirmed this for gpt-5.6-luna, gpt-5.6-sol, and gpt-5.6-terra. The Responses error is:

Unsupported value: 'minimal' is not supported with the 'gpt-5.6-sol' model. Supported values are: 'none', 'low', 'medium', 'high', 'xhigh', and 'max'.

Pydantic AI's unified thinking contract says unsupported levels map to the closest available value, so thinking='minimal' should map to low for GPT-5.6. An explicit provider-specific openai_reasoning_effort='minimal' should remain unchanged and let OpenAI report the incompatibility, consistent with provider-specific setting behavior.

Minimal, Reproducible Example

from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIResponsesModel

agent = Agent(OpenAIResponsesModel('gpt-5.6-sol'), model_settings={'thinking': 'minimal'})
agent.run_sync('Reply with OK.')  # OpenAI returns HTTP 400

Python, Pydantic AI & LLM client version

• Python: 3.12.9
• Pydantic AI: current main (051a92a2)
• LLM provider SDK: OpenAI 2.63.0

[pydanty]

pydanty status: bug -- confirmed (medium), signal 7/10 - p:2-high

Why: Users setting the unified thinking='minimal' with GPT-5.6 models get an HTTP 400 error instead of having the value clamped to the closest supported effort ('low'). This violates the documented unified thinking contract and breaks a default-path workflow on a Tier-1 provider. Other providers already clamp (Mistral, Groq); OpenAI is the outlier. Fixing this makes thinking='minimal' work seamlessly across all model families, matching the promise the docs make.
Evidence: Reproducer confirmed from checked-in source: OPENAI_REASONING_EFFORT_MAP in profiles/openai.py:28-36 maps 'minimal'→'minimal' (static 1:1, no clamping). Both _translate_thinking methods (models/openai.py:2734 ResponsesModel, :995 ChatModel) do direct lookups with no model-specific logic. _ReasoningSupport (profiles/openai.py:58-81) has only enabled_by_default/can_be_disabled/supports_mode/supports_context — no field tracking which effort levels each model accepts. GPT-5.6 is listed at line 103. Repositorian confirmed docs at https://ai.pydantic.dev/thinking state "unsupported levels map to the closest available value" with clamping precedent for Mistral (all→'high') and Groq (minimal/low→'low'). DouweM merged unified thinking in PR #4640; no maintainer discussion about OpenAI effort clamping specifically. Explorer could not independently confirm from OpenAI primary docs that GPT-5.6 rejects 'minimal' — JS-rendered pages were not fetchable; the reporter's error message is consistent with OpenAI API error format but the specific GPT-5.6 effort support list is unverified from primary sources.
Next: pr_ready -- Confirmed code-level contract violation: OpenAI adapter maps thinking='minimal' 1:1 to reasoning_effort='minimal' with no clamping, but docs promise "unsupported levels map to the closest available value." Fix is localized — add per-family effort support tracking to profile and clamp in _translate_thinking. Provider-specific openai_reasoning_effort stays pass-through.

@adtyavrdhn -- assigned for prioritization (p:2-high, signal 7/10).

💰 Run usage: 2,433,594 tokens, 109 model requests (est. cost $1.0986)

Agent	Cost	Input	Output	Requests	Tool calls	Model
repositorian	$0.2966	1,067,506	3,315	40	49	kimi-k2.7-code, glm-5.2
reproducer	$0.2588	565,092	5,941	33	51	kimi-k2.7-code, glm-5.2
code_librarian	$0.3670	465,339	33,801	19	40	—
pydanty-orchestrator	$0.1705	249,782	16,951	14	14	—
explorer	$0.0057	25,295	572	3	4	minimax-m3
total	$1.0986	2,373,014	60,580	109	158

[pydanty]

pydanty runtime failure

Run df_run_89fc3c3a818344de96a6 started from the GitHub webhook, but the Dogfooding runtime
failed before pydanty completed the workflow. The pydanty status comment above
(if any) is from a completed run and still stands.

Next step: maintainer attention needed; rerun after fixing the runtime failure.

Failure detail (tail)

a
16:39:56.580     openrouter usage
16:39:56.589     running tool: replace_repository_lines
16:39:56.710     chat fallback:glm-5.2,gpt-5.6-luna
16:40:01.223     openrouter usage
16:40:01.232     running tool: shell
16:40:01.404     chat fallback:glm-5.2,gpt-5.6-luna
gh-token refreshed
16:40:16.223     openrouter usage
16:40:16.229     running tool: shell
16:40:16.365     chat fallback:glm-5.2,gpt-5.6-luna
16:40:43.652     openrouter usage
16:40:43.660     running tool: shell
16:40:43.954     chat fallback:glm-5.2,gpt-5.6-luna
16:40:57.753     openrouter usage
16:40:57.760     running tool: write_plan
16:40:57.768     running tool: shell
16:40:58.050     chat fallback:glm-5.2,gpt-5.6-luna
16:41:54.855     openrouter usage
16:41:54.861     running tool: shell
16:41:55.064     chat fallback:glm-5.2,gpt-5.6-luna
16:42:11.230     openrouter usage
16:42:11.238     running tool: shell
16:42:11.575     chat fallback:glm-5.2,gpt-5.6-luna
16:42:36.960     openrouter usage
16:42:36.966     running tool: shell
16:42:37.738     chat fallback:glm-5.2,gpt-5.6-luna
16:42:40.576     openrouter usage
16:42:40.583     running tool: shell
16:42:40.736     chat fallback:glm-5.2,gpt-5.6-luna
16:42:46.723     openrouter usage
16:42:46.728     running tool: shell
16:42:47.050     chat fallback:glm-5.2,gpt-5.6-luna
16:42:49.940     openrouter usage
AGENT_ERROR (relevant leg):
  File "/home/pydanty/dogfooding/.runner.py", line 204, in <module>
    asyncio.run(main())
  File "/home/pydanty/dogfooding/.runner.py", line 169, in main
    result = await result
             ^^^^^^^^^^^^
UsageLimitExceeded: Exceeded the input_tokens_limit of 3000000 (input_tokens=3006612). Consider raising the limit, or see the docs on usage limits for budget-aware patterns: https://ai.pydantic.dev/agent/#usage-limits

[daikeren]

I have a tested fix ready in #7082. The approach is:

• add a GPT-5.6 model-profile flag for the unified minimal-effort fallback
• map thinking="minimal" to "low" in both Chat Completions and Responses
• keep explicit openai_reasoning_effort="minimal" settings unchanged
• cover both flag paths, all GPT-5.6 profiles, and a recorded Responses request

Would you be happy with this approach and with me championing the fix? If so, could you assign #7081 to me and confirm whether I should reopen #7082?

thanks!