pydantic-ai
Coverage, reference pages, tools, and guides connected to this topic.
-
pydantic-ai 1.102.0 released
Pydantic AI 1.102.0 fixes a security issue in URL validation by expanding IPv6 transition-form handling to block an SSRF cloud-metadata bypass, but it only matters if you explicitly use `FileUrl(force_download='allow-local')` with untrusted URLs on NAT64/ISATAP networks. It also adds support for additional IPv6 transition forms in URL validation, which makes agent file-download and URL-handling flows more robust against edge-case network inputs.
-
pydantic-ai 1.100.0 released
Pydantic AI 1.100.0 is mainly a security release for agent builders: it normalizes IPv6 transition forms in URL validation to close an SSRF cloud-metadata blocklist bypass affecting `FileUrl(..., force_download='allow-local')` when the URL can be influenced by untrusted input. It does not appear to add major new agent capabilities or performance improvements in the excerpt, and the fix was already shipped in v1.99.0.
-
pydantic-ai 1.99.0 released
Pydantic AI 1.99.0 adds support for the `gemini-3.5-flash` model, giving agent builders another model option to use in production workflows. It also fixes an OpenAI strict-schema bug involving regex lookarounds, which should improve reliability when validating or generating structured outputs.
-
pydantic-ai 1.98.0 released
pydantic-ai 1.98.0 adds OpenAI Responses input token counting via `OpenAIResponsesModel.count_tokens`, which helps agent builders estimate usage/costs more accurately before sending requests. It also replaces the old `Agent(tool_retries=..., output_retries=...)` knobs with a single `retries: int | AgentRetries` setting, and deprecates `pydantic_ai.ext.aci` (`tool_from_aci` and `ACIToolset`) as part of V2 preparation.
-
pydantic-ai 1.97.0 released
Pydantic AI 1.97.0 adds `OnlineEvaluator.run_on_errors` so evals can now run on failed calls, and it changes Google provider setup by splitting `GoogleProvider(vertexai=True|False)` into `GoogleProvider` plus `GoogleCloudProvider` while renaming provider IDs from `google-gla:` to `google:` and `google-vertex:` to `google-cloud:`. It also introduces `MCPToolset` based on `fastmcp-slim[client]` and deprecates `MCPServer*` and `FastMCPToolset`, while setting `ModelResponse.state` to `incomplete` during streaming so builders can more accurately track in-progress responses.
-
pydantic-ai 1.96.0 released
Pydantic-ai 1.96.0 introduces V2 preparation changes, including the new `openai-chat:` prefix with warnings for bare `openai:` (which will switch to Responses API in v2), deprecation of `Agent(history_processors=)` in favor of `capabilities=[ProcessHistory(...)]`, and deprecation of method-style accessors like `result.usage()` for property-style equivalents. It also deprecates `AGUIApp`, `Agent.to_ag_ui()`, and related shims in favor of `AGUIAdapter`, signaling upcoming shifts for AI agent builders toward more standardized APIs.
-
pydantic-ai 1.95.0 released
Pydantic AI 1.95.0 introduces native Tool Search for Anthropic and OpenAI with custom strategies on any provider, an Instrumentation capability replacing Agent(instrument=...), and structured output plus tool combination support for Gemini 3. It prepares for V2 by renaming “built-in tools” to “native tools”, deprecating old fields, and registering them via capabilities=[NativeTool(...)].
-
pydantic-ai 1.94.0 released
Pydantic-ai 1.94.0 adds support for OpenAI's multiple system messages capability through a new profile flag, enabling more flexible prompt structuring for OpenAI models. The release also removes `mistralai` as a direct dependency, streamlining the package's dependency footprint.
-
pydantic-ai 1.93.0 released
Pydantic AI 1.93.0 adds a `tool_choice` setting for more control over model tool selection and improves event handling by yielding `OutputToolCallEvent`/`OutputToolResultEvent` for output tool calls while deprecating function-tool events for failing cases. The release also fixes a bug where spawned tasks weren't properly drained during agent cancellation, improving reliability in concurrent scenarios.
-
pydantic-ai 1.92.0 released
Pydantic AI 1.92.0 introduces Anthropic task budget support and runtime `output_retries` override with deprecation of the old `retries` field, enhancing control over AI agent execution and reliability. It also fixes key bugs like streaming response cleanup on cancellation, MCP session task isolation to prevent exit scope errors, and proper population of `RunContext` with run/conversation IDs and metadata.