Skip to content

fix: Add gpt-5 models to reasoning model detection for temperature parameter handling #2483

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 3 commits into from
Aug 11, 2025
Merged

fix: Add gpt-5 models to reasoning model detection for temperature parameter handling #2483

merged 3 commits into from
Aug 11, 2025
+87 −24

Conversation

@DouweM
Copy link
Collaborator

@DouweM DouweM commented Aug 8, 2025

Extended openai_model_profile() to detect gpt-5 models as reasoning models that don't support sampling settings like temperature.

Changes:

  • Updated reasoning model detection logic to include models starting with 'gpt-5'
  • GPT-5 models now have openai_supports_sampling_settings=False
  • Prevents temperature, presence_penalty, frequency_penalty from being sent to GPT-5 models
  • Resolves API errors when temperature parameter is provided with gpt-5 models

Fixes #2482

Generated with Claude Code

@claude @DouweM
fix: Add gpt-5 models to reasoning model detection for temperature pa…
1d91ca1 
…rameter handling

- Extended openai_model_profile() to detect gpt-5 models as reasoning models
- GPT-5 models now have openai_supports_sampling_settings=False to prevent
  temperature, presence_penalty, frequency_penalty, etc. from being sent
- Resolves issue where temperature parameter caused API errors with gpt-5 models
- Maintains existing logic for o-series models

Fixes #2482

Co-authored-by: Douwe Maan <DouweM@users.noreply.github.com>
@DouweM DouweM self-assigned this Aug 8, 2025
@github-actions
Copy link

github-actions bot commented Aug 8, 2025 •
edited
Loading

Docs Preview

commit: 6ec049c
Preview URL: https://ea1fd6d2-pydantic-ai-previews.pydantic.workers.dev

@Kludex Kludex enabled auto-merge (squash) August 11, 2025 08:50
@Kludex Kludex merged commit bc9a2fd into main Aug 11, 2025
16 checks passed
@Kludex Kludex deleted the claude/issue-2482-20250808-1310 branch August 11, 2025 08:58
ethanabrooks added a commit to reflectionai/pydantic-ai that referenced this pull request Aug 20, 2025
@cvanderschuere @akenarsari
@tonyxwz @DouweM @ethan01x @xjose97x @Kludex @strawgate @dmontagu @GuillermoBlasco @HamzaFarhan @medaminezghal @Tiksagol @mattbrandman @claude @AlexEnrique @jerry-heygen @claude @spike-spiegel-21 @alexmojaki @jerry-reevo @ryx2 @kauabh @Viicos @ethanabrooks @eballesteros
Update to v0.7.2 (#9)
f5a0a15 
* Add `priority` `service_tier` to `OpenAIModelSettings` and respect it in `OpenAIResponsesModel` (pydantic#2368)

* Add an example of using RunContext to pass data among tools (pydantic#2316)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Rename gemini-2.5-flash-lite-preview-06-17 to gemini-2.5-flash-lite as it's out of preview (pydantic#2387)

* Fix toggleable toolset example so toolset state is not shared across agent runs (pydantic#2396)

* Support custom thinking tags specified on the model profile (pydantic#2364)

Co-authored-by: jescudero <jescudero@itos.es>
Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Add convenience functions to handle AG-UI requests with request-specific deps (pydantic#2397)

* docs: add missing optional packages in `install.md` (pydantic#2412)

* Include default values in tool arguments JSON schema (pydantic#2418)

* Fix "test_download_item_no_content_type test fails on macOS" (pydantic#2404)

* Allow string format, pattern and others in OpenAI strict JSON mode (pydantic#2420)

* Let more `BaseModel`s use OpenAI strict JSON mode by defaulting to `additionalProperties=False` (pydantic#2419)

* BREAKING CHANGE: Change type of 'source' field on EvaluationResult (pydantic#2388)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Fix ImageUrl, VideoUrl, AudioUrl and DocumentUrl not being serializable (pydantic#2422)

* BREAKING CHANGE: Support printing reasons in the console output for pydantic-evals (pydantic#2163)

* Document performance implications of async vs sync tools (pydantic#2298)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Mention that tools become toolset internally (pydantic#2395)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Fix tests for Logfire>=3.22.0 (pydantic#2346)

* tests: speed up the test suite (pydantic#2414)

* google: add more information about schema on union (pydantic#2426)

* typo in output docs (pydantic#2427)

* Deprecate `GeminiModel` in favor of `GoogleModel` (pydantic#2416)

* Use `httpx` on `GoogleProvider` (pydantic#2438)

* Remove older deprecated models and add new model of Anthropic (pydantic#2435)

* Remove `next()` method from `Graph` (pydantic#2440)

* BREAKING CHANGE: Remove `data` from `FinalResult` (pydantic#2443)

* BREAKING CHANGE: Remove `get_data` and `validate_structured_result` from `StreamedRunResult` (pydantic#2445)

* docs: add `griffe_warnings_deprecated` (pydantic#2444)

* BREAKING CHANGE: Remove `format_as_xml` module (pydantic#2446)

* BREAKING CHANGE: Remove `result_type` parameter and similar from `Agent` (pydantic#2441)

* Deprecate `GoogleGLAProvider` and `GoogleVertexProvider` (pydantic#2450)

* BREAKING CHANGE: drop 4 months old deprecation warnings (pydantic#2451)

* Automatically use OpenAI strict mode for strict-compatible native output types (pydantic#2447)

* Make `InlineDefsJsonSchemaTransformer` public (pydantic#2455)

* Send `ThinkingPart`s back to Anthropic used through Bedrock (pydantic#2454)

* Bump boto3 to support `AWS_BEARER_TOKEN_BEDROCK` API key env var (pydantic#2456)

* Add new Heroku models (pydantic#2459)

* Add `builtin_tools` to `Agent` (pydantic#2102)

Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>
Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Bump mcp-run-python (pydantic#2470)

* Remove fail_under from top-level coverage config so <100% html-coverage step doesn't end CI run (pydantic#2475)

* Add AbstractAgent, WrapperAgent, Agent.event_stream_handler, Toolset.id, Agent.override(tools=...) in preparation for Temporal (pydantic#2458)

* Let toolsets be built dynamically based on run context (pydantic#2366)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Add ToolsetFunc to API docs (fix CI) (pydantic#2486)

* tests: change time of evals example (pydantic#2501)

* ci: remove html and xml reports (pydantic#2491)

* fix: Add gpt-5 models to reasoning model detection for temperature parameter handling (pydantic#2483)

Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: Douwe Maan <DouweM@users.noreply.github.com>
Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>

* History processor replaces message history (pydantic#2324)

Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>

* ci: split test suite (pydantic#2436)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* ci: use the right install command (pydantic#2506)

* Update config.yaml (pydantic#2514)

* Skip testing flaky evals example (pydantic#2518)

* Fix error when parsing usage details for video without audio track in Google models (pydantic#2507)

* Make OpenAIResponsesModelSettings.openai_builtin_tools work again (pydantic#2520)

* Let Agent be run in a Temporal workflow by moving model requests, tool calls, and MCP to Temporal activities (pydantic#2225)

* Install only dev in CI (pydantic#2523)

* Improve CLAUDE.md (pydantic#2524)

* Add best practices regarding to coverage to CLAUDE.md (pydantic#2527)

* Add support for `"openai-responses"` model inference string (pydantic#2528)

Co-authored-by: Claude <noreply@anthropic.com>

* docs: Confident AI (pydantic#2529)

* chore: mention what to do with the documentation when deprecating a class (pydantic#2530)

* chore: drop hyperlint (pydantic#2531)

* ci: improve matrix readability (pydantic#2532)

* Add pip to dev deps for PyCharm (pydantic#2533)

Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>

* Add genai-prices to dev deps and a basic test (pydantic#2537)

* Add `--durations=100` to all pytest calls in CI (pydantic#2534)

* Cleanup snapshot in test_evaluate_async_logfire (pydantic#2538)

* Make some minor tweaks to the temporal docs (pydantic#2522)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Add new OpenAI GPT-5 models (pydantic#2503)

* Fix `FallbackModel` to respect each model's model settings (pydantic#2540)

* Add support for OpenAI verbosity parameter in Responses API (pydantic#2493)

Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* Add `UsageLimits.count_tokens_before_request` using Gemini `count_tokens` API (pydantic#2137)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* chore: Fix uv.lock (pydantic#2546)

* Stop calling MCP server `get_tools` ahead of `agent run` span (pydantic#2545)

* Disable instrumentation by default in tests (pydantic#2535)

Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>

* Only wrap necessary parts of type aliases in forward annotations (pydantic#2548)

* Remove anthropic-beta default header set in `AnthropicModel` (pydantic#2544)

Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>

* docs: Clarify why AG-UI example links are on localhost (pydantic#2549)

* chore: Fix path to agent class in CLAUDE.md (pydantic#2550)

* Ignore leading whitespace when streaming from Qwen or DeepSeek (pydantic#2554)

* Ask model to try again if it produced a response without text or tool calls, only thinking (pydantic#2556)

Co-authored-by: Douwe Maan <douwe@pydantic.dev>

* chore: Improve Temporal test to check trace as tree instead of list (pydantic#2559)

* Fix: Forward max_uses parameter to Anthropic WebSearchTool (pydantic#2561)

* Let message history end on ModelResponse and execute pending tool calls (pydantic#2562)

* Fix type issues

* skip tests requiring API keys

* add `google-genai` dependency

* add other provider deps

* add pragma: no cover for untested logic

---------

Co-authored-by: akenar <52220260+akenarsari@users.noreply.github.com>
Co-authored-by: Tony Woland <16152581+tonyxwz@users.noreply.github.com>
Co-authored-by: Douwe Maan <douwe@pydantic.dev>
Co-authored-by: Yi-Chen Lin <103916325+ethan01x@users.noreply.github.com>
Co-authored-by: José I. Escudero <joseignacioescudero@gmail.com>
Co-authored-by: jescudero <jescudero@itos.es>
Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>
Co-authored-by: William Easton <bill.easton@elastic.co>
Co-authored-by: David Montague <35119617+dmontagu@users.noreply.github.com>
Co-authored-by: Guillermo <guillermo@mankind.technology>
Co-authored-by: Hamza Farhan <thehamza96@gmail.com>
Co-authored-by: Mohamed Amine Zghal <medaminezghal@outlook.com>
Co-authored-by: Yinon Ehrlich <Tiksagol@users.noreply.github.com>
Co-authored-by: Matthew Brandman <matthb6@gmail.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: Douwe Maan <DouweM@users.noreply.github.com>
Co-authored-by: Alex Enrique <41076109+AlexEnrique@users.noreply.github.com>
Co-authored-by: Jerry Yan <jerry@heygen.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Mayank <83648453+spike-spiegel-21@users.noreply.github.com>
Co-authored-by: Alex Hall <alex.mojaki@gmail.com>
Co-authored-by: Jerry Lin <jerry@reevo.ai>
Co-authored-by: Raymond Xu <raymond.y.xu@gmail.com>
Co-authored-by: kauabh <56749351+kauabh@users.noreply.github.com>
Co-authored-by: Victorien <65306057+Viicos@users.noreply.github.com>
Co-authored-by: Ethan Brooks <ethanabrooks@gmail.com>
Co-authored-by: eballesteros <44843469+eballesteros@users.noreply.github.com>
@matthewfranglen matthewfranglen mentioned this pull request Oct 13, 2025
Open
5 tasks
matthewfranglen added a commit to matthewfranglen/pydantic-ai that referenced this pull request Oct 13, 2025
@matthewfranglen
FunctionTextFormat -> TextFormat, change handling
3927cf0 
This separates the implementation of regex and lark to two classes.
The base class of these would be empty, so the `TextFormat` has become a
union type.
The openai tool call handling has been changed to silently ignore
formats that it does not handle. This is consistent with how the gpt-5
models ignore temperature parameter, which is not supported by reasoning
models (see pydantic#2483).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Kludex Kludex Kludex approved these changes

+1 more reviewer

@ZinoKader ZinoKader ZinoKader approved these changes

Reviewers whose approvals may not affect merge requirements

Assignees

@DouweM DouweM

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

GPT-5 fail gracefully if temperature parameter is provided

4 participants

@DouweM @ZinoKader @Kludex