Pydantic AI

What it is for

Pydantic AI is a typed Python agent framework: you declare an Agent, tools, and a Pydantic output type, then run it against any supported model. Use it when the job is write the agent in Python, not when you need a hosted AgentOS control plane.

The framework is MIT open source (pydantic/pydantic-ai). The library itself is $0. Official observability is Pydantic Logfire: Personal is free; Team is $49/month.

What makes it unique: the cover is Pydantic’s official Logfire plan grid — Personal Free, Team $49/mo, Growth $249/mo, 10M records and $2/M overage on the same cards.

Key features

• Typed agents — Agent("openai:…", output_type=MyModel) with Pydantic validation on the structured result (pydantic-ai).
• Function tools — @agent.tool / @agent.tool_plain generate JSON schemas from type hints (pydantic-ai).
• Model-agnostic providers — OpenAI, Anthropic, Gemini, Bedrock, Groq, Ollama, Nebius, and others behind one Agent string (docs).
• Durable execution — Temporal, DBOS, Prefect, Restate, or AWS Lambda so a long run survives a crash (pydantic-ai).
• Streaming to a UI — AG-UI and Vercel AI Data Stream protocols for token / tool events (pydantic-ai).
• Evals + Logfire traces — datasets and scores in Pydantic Evals; every run, model call, and tool span in Logfire (Logfire).

Who it's for

Best fit: Python teams who already trust Pydantic and will add Logfire when they want a trace.

Poor fit: teams that want a hosted AgentOS UI without writing the agent (Agno).

Pricing

Verified 2026-10-22 (Asia/Shanghai) from pydantic.dev/pricing (cover) and pydantic-ai. The framework is free. Dollars below are Logfire.

Pydantic AI (library) — $0. MIT. pip install pydantic-ai / uv add pydantic-ai. You still pay the model vendor.

Logfire (official observability)

• Personal — Free, forever. 10 million telemetry records / month, hard-capped at $0. 3 projects. 30-day retention. 1 seat + 2 read-only guests. No card.
• Team — $49/month. 10M records included, then $2 / million. 5 seats included ($25 extra). Evals + human review. 30-day retention.
• Growth — $249/month. Everything in Team. Unlimited seats / guests / projects. Up to 90-day retention.
• Enterprise — custom. Dedicated / self-host, SSO / SCIM, BAA.

Every paid card includes a $20/month record credit (the same 10M). Personal pauses ingestion at the cap so you cannot owe money on that tier.

Limitations / costs

• The $0 library is not a $0 model bill.
• Personal pauses at 10M records. A chatty agent loop will hit it.
• Python only. No Go / TS SDK.
• Not AgentOS. There is no hosted “serve my agent to Slack” control plane here — that is Agno.

Alternatives

• Agno — AgentOS runtime + $150 live control plane if you want to serve the agent.
• LangChain — graphs and a huge integration catalog.
• CrewAI — role-based crews if that is already the abstraction.

Official tutorial

We searched official Pydantic YouTube for a first-party Pydantic AI getting-started video we could oEmbed. Ranked walkthroughs (BugBytes and others) are not first-party, so they are not embedded. Use the written Install plus Agents and Logfire.

Website

https://pydantic.dev/pydantic-ai

Repository

https://github.com/pydantic/pydantic-ai

Source

Catalog row curated for AI Tools Nav. Cover is Pydantic Logfire’s official Personal Free / Team $49 / Growth $249 cards. Framework is MIT $0. No first-party getting-started embed.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.