settings be passed to model classes by @svilupp in #2136StructuredDict for structured outputs with custom JSON schema by @fswair in #2157Agent.iter() by @erosennin in #2168_ensure_decodeable by 634% by @misrasaurabh1 in #2155EvaluationReport and ReportCase into generic dataclasses by @dmontagu in #1799ToolDefinition.description optional and fix Bedrock description handling by @dmontagu in #1507AgentStream.stream_output (available inside agent.iter) stream validated output data instead of raising validation errors by @DouweM in #2134model_request_stream_sync to direct API by @hewliyang in #2116{FunctionToolCallEvent,FunctionToolResultEvent}.call_id in favor of tool_call_id by @proever in #2028RetryPromptPart not tied to a tool call contains validation feedback rather than a user message by @hovi in #2008GoogleProvider by @Kludex in #2031ThinkingPart to preceed TextPart in OpenAIResponsesModel by @Kludex in #2043ThinkingPart back when using tool calls by @Kludex in #2072NativeOutput and PromptedOutput modes in addition to ToolOutput by @DouweM in #1628UserPromptNode by @Kludex in #2052ThinkingPart when delta contains signature by @Kludex in #2012See ###1142 — Adds support for thinking parts.
We now convert the thinking blocks ("<think>..."</think>") in provider specific text parts to
PydanticAI ThinkingParts. Also, as part of this release, we made the choice to not send back the
ThinkingParts to the provider - the idea is to save costs on behalf of the user. In the future, we
intend to add a setting to customize this behavior.
McpError from MCP tool calls by @ppcantidio in #1999process_tool_call hook to MCP servers to modify tool args, metadata, and return value by @stevenh in #2000ModelSettings.timeout in GoogleModel by @DouweM in #2006RunContext support to history processors by @Wh1isper in #2002max_tokens to 4096 by default by @Kludex in #1994history_processors parameter to Agent for message processing by @Kludex in #1970GoogleModel instead of GeminiModel on inference by @Kludex in #1881InstrumentedModel by @alexmojaki in #1898service_tier to OpenAIModelSettings by @empezarcero in #1923httpx.AsyncClient in MCPServerHTTP by @mpfaffenberger in #1925temperature and top_p to OpenAI reasoning models by @DouweM in #1956fileData (direct file URL) for GeminiModel and GoogleModel by @vricciardulli in #1136HerokuProvider by @Kludex in #1933stop_sequences to Google models by @Kludex in #1935vendor_id and vendor_details.finish_reason to Gemini/Google model responses by @davide-andreoli in #1800sse_read_timeout timedelta by @alexmojaki in #1843output_type, as well as lists of functions and other types by @DouweM in #1785ModelProfile to let model-specific behaviors be configured independent of the model class by @DouweM in #1835mcp>=1.9.0 on the pydantic-ai-slim[mcp] extra by @DouweM in #1840Agent.to_cli to str by @barp in #1823labels for GeminiModel and GoogleModel on Vertex AI by @vricciardulli in #1056n parameter from OpenAIModel requests by @kiqaps in #1807tool_prefix option to MCP servers and error on conflicting tool names by @Wh1isper in #1266RunContext to not be documented when require_parameter_descriptions=True as it's not passed to the model anyway by @Kevsnz in #1750prog_name customizable by @fswair in #1804AsyncBeta instead of AsyncAnthropic by @Kludex in #1818prepare_tools param to Agent class by @rmaceissoft in #1474GoogleModel by @Kludex in #1780prepare_tools to _prepare_tools by @Kludex in #1789Agent.to_cli() by @Kludex in #1791'openrouter' string to OpenAIModel provider param by @Kludex in #1792capabilities required on AgentCard by @Kludex in #1733include_binary_content flag to InstrumentationSettings, rename OTel attribute key from content to binary_content for BinaryParts by @alexmojaki in #1739vendor_id to the model response. by @peterHoburg in #1547clai should not require OPENAI_API_KEY by @samuelcolvin in #1724sys.path by @Kludex in #1703include_metadata and include_expected_output by @lionpeloux in #1700direct public API by @samuelcolvin in #1599to_cli() method to Agent by @AndrewHannigan in #1642~/.pydantic-ai by @Kludex in #1701BinaryContent information by @Kludex in #1603See ###1647 — usage makes sense as part of ModelResponse, and could be really useful in "messages" (really a sequence of requests and response). In this PR:
usage to ModelResponse (field has a default factory of Usage() so it'll work to load data that doesn't have usage)Model.request to just ModelResponse instead of tuple[ModelResponse, Usage]ToolCallPartDelta.tool_call_id when known from previous delta by @DouweM in #1694clai by @samuelcolvin in #1504temperature to 0 on BedrockConverseModel by @pandersen-uncharted in #1631extra_headers to ModelSettings by @ramanan-ravi in #1644thinking_config to GeminiModel by @Kludex in #1609HandleResponseEvent by @DouweM in #1619base_url to Mistral Provider by @alibeyram in #1617use_attribute_docstrings=True by default on tools by @Kludex in #1605report_case_name in span by @alexmojaki in #1544list type inside the Annotated on function_declarations field by @gitraffe in #1570prepare_env.py to Python file in mcp-run-python by @samuelcolvin in #1461instructions by @Kludex in #1529extra_body to ModelSettings by @Kludex in #1538BinaryContent in message_history by @arde100 in #1523BinaryContent by decoding base64 by @DouweM in #1513StdioServerParameters.cwd param by @timesler in #1514UserPromptNode by @Kludex in #1503additionalProperties by gemini by @Kludex in #1506BedrockConverseModel by @leandrodamascena in #1435result -> output by @samuelcolvin in #1248BedrockConverseModel): add additional configuration fields to Bedrock Runtime API by @leandrodamascena in #1458gemini-2.5-pro-preview-03-25 model (paid version of gemini 2.5 pro) by @barapa in #1473format_as_xml by @samuelcolvin in #1484instructions parameter by @Kludex in #1360user_prompt optional by @Kludex in #1406stop_sequences to ModelSettings by @Kludex in #1419handle_stream.stream_output() in the CLI by @Kludex in #1376mcp-run-python server to use deno by @samuelcolvin in #1340exclusiveMaximum/exclusiveMinimum from Gemini by @Kludex in #1341generate_summary and truncation to OpenAIResponsesModelSettings by @Kludex in #1328ImageUrl and DocumentUrl for Anthropic Models by @jnishiyama in #1318pydantic-evals package by @dmontagu in #935PYTHONOPTIMIZE=1 by @Kludex in #1307headers, timeout, and sse_read_timeout to MCPServerHTTP by @JohnUiterwyk in #1218get_running_loop instead of get_event_loop" by @Kludex in #1252user mapping in openai chat completion by @ovisek in #1174max_tokens to max_completion_tokens by @barapa in #1206model_request_parameters attribute (containing tool definitions) to chat spans by @alexmojaki in #1177get_running_loop instead of get_event_loop by @Kludex in #1204pydantic-ai-slim version on CLI instead of pydantic-ai by @samuelcolvin in #1213system parameter from OpenAIModel by @Kludex in #1235SystemPromptPart by @vricciardulli in #1154MCPServerSSE to MCPServerHTTP and more docs by @samuelcolvin in #1176OpenAIProvider to use environment variable for base URL by @hrahmadi71 in #1117AzureProvider by @Kludex in #1091InstrumentedModel and FallbackModel fixes by @alexmojaki in #1121DocumentUrl and support document via BinaryContent by @Kludex in #987base_url to models, populate server.address and server.port in spans by @alexmojaki in #1074This release only makes sure that pydantic-ai-examples have the right dependency versions.
Agent.instrument_all() method to instrument all agents by default by @alexmojaki in #1047pai CLI by @Kludex in #1031instrument param to Agent, use plain OpenTelemetry by @alexmojaki in #1005gemini-2.0-pro-exp-02-05 by @barapa in #1038return_type by @Kludex in #1017model request span with InstrumentedModel by @alexmojaki in #1012all_messages in agent span with all_messages_events in same format as InstrumentedModel span by @alexmojaki in #1018.iter() API to fully replace existing streaming implementation by @dmontagu in #951TestModel by @Kludex in #1015max_results parameter to DuckDuckGo search tool by @HamzaFarhan in #1003ImageUrl by @pedroallenrevez in #993FallbackModel support by @sydney-runkle in #894InstrumentedModel by @alexmojaki in #892name methods from openai and mistral models by @sydney-runkle in #925request_stream to InstrumentedModel by @alexmojaki in #922next more ergonomic by @dmontagu in #833InstrumentedModel by @alexmojaki in #945async from _get_model by @alexmojaki in #893ModelResponse.model_name from responses by @alexmojaki in #883o3 support for OpenAIModel by @sydney-runkle in #853AgentModel class by @dmontagu in #800reasoning_effort param for OpenAIModel by @sydney-runkle in #854project_id == creds_id check for vertexai by @sydney-runkle in #862model_name and system model properties by @sydney-runkle in #865pydantic_ai.Agent to using pydantic_graph by @dmontagu in #725tool_call_id to chatbot.append if it's set by @chug2k in #771AgentDeps refs to AgentDepsT by @zzstoatzz in #767ModelSettings to support specialized model requests by @sydney-runkle in #766_utils.Either by @dmontagu in #768docstring_format and require_parameter_descriptions by @chris-dare in #776OllamaModel in favor of usage with OpenAIModel by @sydney-runkle in #805Cohere docs and additional (live) tests by @sydney-runkle in #810allow_model_requests check by @YanSte in #697model_name to ModelResponse by @sydney-runkle in #701'deepseek-r1' to the list of Ollama Model Names by @izzyacademy in #735o1 by @sydney-runkle in #740from_text and from_tool_call utilities that complicate snapshot testing by @sydney-runkle in #744IsNow() delta by @dmontagu in #746set_event_loop fixture by @dmontagu in #747parallel_tool_calls in ModelSettings by @sydney-runkle in #750user role system prompts o1-preview-2024-09-12 by @siavashg in #754asyncio to all examples that need it. by @ME-Msc in #638docstring_format and require_parameter_descriptions on tools by @sydney-runkle in #643phi4 to Ollama by @josead in #668run/ run_sync docs by @KranthiGV in #688testing-evals and RAG example by @lealre in #694UV_FROZEN globally in CI by @samuelcolvin in #626dynamic to system_prompt decorator, allowing reevaluation by @josead in #560--reinstall instead of uninstall/install for docs insiders packages by @T-256 in #627result_type on a run by @samuelcolvin in #629scope to google.auth.default() by @jtbaker in #579AgentDeps default to None. by @samuelcolvin in #592X | None = None with Gemini by @samuelcolvin in #540chat_app example by @DurandA in #553RunContext by @samuelcolvin in #570capture_run_messages support nested agent calls by @samuelcolvin in #573TestModel handles result retries correctly by @jlowin in #572ResultData, some related cleanup by @samuelcolvin in #512Agents page by @lealre in #521last_run_messages, add capture_run_messages by @samuelcolvin in #536ModelMessage discriminator in chat_app example by @DurandA in #402Cost to Usage by @dmontagu in #403openai:o1 model support by @sydney-runkle in #498RunContext more widely by @samuelcolvin in #500All releases prior to V1 can contain breaking changes according to semvar, but this release contains more than usual:
ToolReturnPart for each tool call, see #184 and #274tests/test_tools.py by @samuelcolvin in #229uv version by @zzstoatzz in #236ModelSettings logic by @sydney-runkle in #227messages.py by @sydney-runkle in #255defer_build with type adapters, not custom _LazyTypeAdapter by @sydney-runkle in #253isinstance checks for message kinds by @sydney-runkle in #252messages to RunContext by @samuelcolvin in #257list[ModelRequest | ModelResponse] by @samuelcolvin in #259.stream_text(delta=True) by @samuelcolvin in #178IndexError when streaming OpenAI by @samuelcolvin in #181Agent.name by @samuelcolvin in #141bank_support.py -- fix docstring typo by @akgerber in #143logfire optional group by @samuelcolvin in #154nee -> need by @patrick91 in #111.vscode/ to .gitignore by @zzstoatzz in #113Agent(tools_[...]) by @samuelcolvin in #128retreiver to tool by @samuelcolvin in #96CallContext -> RunContext by @samuelcolvin in #101pydantic-ai-slim only by @samuelcolvin in #103test-live by @samuelcolvin in #90infer_model by @samuelcolvin in #89pydantic-ai-slim by @samuelcolvin in #92retriever_context -> retriever by @samuelcolvin in #68shared.py by @samuelcolvin in #29ToolCall to Structured in most places by @samuelcolvin in #34ResultData, rename method/attributes appropriately by @samuelcolvin in #36Either from _handle_model_response and _handle_streamed_model_response by @samuelcolvin in #39Agent initialization to take deps type, not deps by @samuelcolvin in #40TypeAliasType unions by @samuelcolvin in #26initial release 🎉