What's Changed

• Add support for aspect ratio in Gemini image generation by @mwildehahn and @ajac-zero in #3672
• Pass container_id back to Anthropic API by @mattbrandman in #3637
• Upgrade temporalio to 1.20.0 by @akrivka in #3679
• Fix error when OpenRouter response includes file annotations by @dsfaccini in #3671
• Suppress broken resource errors if cancelling by @dmontagu in #3675
• Fix GoogleModel thinking signature not stored on tool calls when streaming by @jerry-heygen in #3647
• Don't require anthropic dependency when using Anthropic model with other provider by @dsfaccini in #3652

New Contributors

• @akrivka made their first contribution in #3679

Full Changelog: v1.28.0...v1.29.0

What's Changed

• Adds native structured output support for claude-haiku-4-5 by @dsfaccini in #3670
• Support us-gov. and other multi-character Bedrock geo prefixes by @Leundai in #3645
• Allow missing content in streamed OpenRouter reasoning details by @ajac-zero in #3674
• Remove typings stubs by @dsfaccini in #3673

New Contributors

• @Leundai made their first contribution in #3645

Full Changelog: v1.27.0...v1.28.0

What's Changed

• feat: Allow dynamic configuration of built-in tools via RunContext by @AlanPonnachan in #3600
• Support tool and resource caching for MCP servers that support change notifications by @dsfaccini in #3560
• Support raw CoT reasoning from LM Studio and other OpenAI Responses-compatible APIs by @dsfaccini in #3559
• Fix structured output with nested definitions with Gemini via OpenRouter by @dsfaccini in #3618
• fix: Copy extra_headers to allow model_settings reuse by @Wh1isper in #3628
• Add VercelAIAdapter.dump_messages to convert Pydantic AI messages to Vercel AI messages by @dsfaccini in #3392
• Disable use of tool_choice=required for deepseek-reasoner by @ZeroAurora in #3644
• Use custom reasoning field for OpenRouter by @xcpky in #3626

New Contributors

• @AlanPonnachan made their first contribution in #3600

Full Changelog: v1.26.0...v1.27.0

What's Changed

• Support JSON object output for Deepseek provider by @xcpky in #3601
• Clarify FallbackModel behavior on ValidationErrors by @dsfaccini in #3603
• Add latest Grok (xAI) models by @darjeeling in #3613
• Automatically omit TTL from cache_control when AnthropicModel is used with Bedrock client by @Wh1isper in #3616
• Add custom reasoning field support to OpenAI model profiles by @ZeroAurora in #3536
• Add gateway/...:... to Known Model Names by @dsfaccini in #3593

New Contributors

• @darjeeling made their first contribution in #3613
• @ZeroAurora made their first contribution in #3536

Full Changelog: v1.25.1...v1.26.0

What's Changed

• Make FastMCPToolset work with DBOS by @qianl15 in #3540
• Fix OpenRouterModel error when native_finish_reason is None by @reuben-if in #3584
• Fix error when OpenRouter API returns None tool call arguments by @barp in #3571

New Contributors

• @reuben-if made their first contribution in #3584

Full Changelog: v1.25.0...v1.25.1

What's Changed

• Support gemini-3-pro-image-preview with Nano Banana Pro by @DouweM in #3576
• Return tool error to Google in 'error' key by @DouweM in #3583

Full Changelog: v1.24.0...v1.25.0

What's Changed

• Support native JSON output and strict tool calls for Anthropic by @dsfaccini in #3457
• Support Pydantic validation context by @NicolasPllr1 in #3448
• Support logprobs output from Responses API by @tim-becker in #3535
• Support instructions-only agent run with OpenAIResponsesModel by @Cody-learns in #3470
• Enable PLW1514 rule (use utf-8 encoding) by @Danipulok in #3524
• Replace direct _model_name access with model_name property in OpenAI models and streamed responses by @BBYNAI in #3564
• Update docs on multi-modal file URLs being downloaded or sent directly by @marmor7 in #3492

New Contributors

• @Danipulok made their first contribution in #3524
• @Cody-learns made their first contribution in #3470
• @marmor7 made their first contribution in #3492
• @tim-becker made their first contribution in #3535

Full Changelog: v1.23.0...v1.24.0

What's Changed

• Add Anthropic built-in WebFetchTool support by @sarth6 in #3427
• Allow user_prompt in HITL by @Wh1isper in #3528
• Add anthropic_cache_messages model setting and automatically strip cache points over the limit by @Wh1isper in #3442
• Add Gemini 3 Pro support to OpenRouterModel by @zeyuyuyu in #3548
• Ensure openrouter_reasoning model setting is sent to API by @xcpky in #3545
• Don't close custom httpx client provided to MCPServerHTTP by @DouweM in #3534
• Fix double counting of request tokens in evals by @dmontagu in #3553
• Fix pydantic_graph.beta.GraphRun GeneratorExit handling by @michgur in #3525
• Add test to ensure we allow message history starting with assistant message (model response) by @restless in #3519
• Remove README.md from wheel for pydantic-ai by @johnslavik in #3521
• Upgrade Ruff to v0.14.6 by @Viicos in #3547
• Use model_name property in OpenAIChatModel chat completion create request by @BBYNAI in #3543
• Pin huggingface-hub to <1 by @Kludex in #3557
• Add docs examples to use the Gateway aside from the gateway/<provider>:<model> shorthand by @dsfaccini in #3484

New Contributors

• @restless made their first contribution in #3519
• @johnslavik made their first contribution in #3521
• @michgur made their first contribution in #3525
• @zeyuyuyu made their first contribution in #3548
• @BBYNAI made their first contribution in #3543
• @sarth6 made their first contribution in #3427

Full Changelog: v1.22.0...v1.23.0

What's Changed

• Add OpenRouterModel as OpenAIChatModel subclass with additional feature support by @ajac-zero in #3089
• Make FallbackModel fall back on all model API errors, not just HTTP status 400+ by @dsfaccini in #3494
• Don't skip model request when history ends on response but there are new instructions by @DouweM in #3509
• Revert "feat: enforce message history starts with user message" by @DouweM in #3508
• Don't send Google messages with 0 parts by @DouweM in #3511
• Immediately raise error when response is empty because of token limit by @DouweM in #3512
• Relax UserError into a warning when state deps is not provided with AG-UI by @MikeRyanDev in #3510
• Add type stubs for some third-party libraries by @lars20070 in #3443
• Add document to allowed cacheable_types for anthropic by @marpulli in #3513
• Don't insert empty ThinkingPart when Google response ends in text with thought_signature by @DouweM in #3516
• Fix deprecated GeminiModel structured output and tool call thought signatures by @DouweM in #3517

New Contributors

• @marpulli made their first contribution in #3513

Full Changelog: v1.21.0...v1.22.0

What's Changed

• Feature: MCP client Resources support by @fennb in #3024
• Expose MCP server instructions in MCPServer.instructions property by @mochow13 in #3431
• Add BinaryContent.from_path convenience method by @dsfaccini in #3482
• Enforce message history starts with user message by @Pavanmanikanta98 in #3440
• Always strip Markdown fences from structured output by @DouweM in #3475
• Fix Gemini nested tool argument schemas by removing title from $defs by @conradlee in #3487
• Fix issue with Gemini and Prefect by removing unused GoogleModel._url by @DouweM in #3499

New Contributors

• @Pavanmanikanta98 made their first contribution in #3440
• @fennb made their first contribution in #3024

Full Changelog: v1.20.0...v1.21.0