Pydantic AI 2.51.0: GPT-Live support and six releases in two weeks
On September 25, 2026, release 2.51.0 of the Pydantic AI platform came out — a Python framework for building AI agents with a typed agent loop. Six releases have accumulated since the last review: 2.46.0, 2.47.0, 2.48.0, 2.49.0, 2.50.0, and 2.51.0.
What's new
- OpenAI GPT-Live support — 2.51.0 adds the OpenAILiveModel model for realtime sessions, plus the context_window_used parameter, which shows context window usage based on provider data.
- DecisionModel for routing — 2.50.0 introduced the DecisionModel base class, which lets an agent choose a route by name, with an optional decision_route_threshold instead of the previous tool-call lean mechanism.
- TypeSafeModel improvements — 2.47.0—2.49.0 expanded type handling: support for None as a route, describing bool field values via BoolCriteria, integers as choice options, and applying default values for nested models.
- More reliable realtime sessions — 2.49.0 added the RealtimeSession.wait_for_reply() method, fixed streaming audio delivery and error handling on the receiving side, and 2.51.0 added buffering of text deltas when replaying completed responses.
- Performance optimization — in 2.51.0 the agent graph is now cached instead of being rebuilt on every Agent.run(), and a task group is no longer created for a single child element in a toolset or capability fan-out.
- New models and providers — 2.48.0—2.50.0 added support for OpenAI gpt-6-sol and gpt-6-luna, Claude Opus 5.5, and the Gemini 3.8 Live realtime models; 2.49.0 introduced GitHubCopilotOAuthFlow for device authorization.
What the platform does
Pydantic AI is a Python SDK with a typed agent loop, where the model is swapped with a single line and the same agent runs behind a web frontend, in a terminal, on a voice call, in a background queue, or in Git. The platform supports multiple agents working together, custom tools and functions, MCP, agent memory, human confirmation, graphs and state persistence, agent quality evaluation, voice and speech, and working with images and documents.
Your own model instead of the cloud
Your own model is connected via Ollama — the only method named in the documentation. The platform is installed with the command uv add pydantic-ai pydantic-ai-harness or uv add pydantic-ai. License — MIT.
How it differs from its neighbors
Among comparable frameworks — CrewAI, Agno, and AgentScope — only Pydantic AI states agent quality evaluation via Pydantic Evals in its documentation. At the same time, the A2A protocol and a knowledge base (RAG), which most of its neighbors have, are not mentioned in the Pydantic AI documentation. Custom tools, MCP support, and agent memory are stated for all four platforms.
| Pydantic AI | CrewAI | Agno | AgentScope | |
|---|---|---|---|---|
| Type | framework | framework | framework | framework |
| Language | Python | Python | Python | Python |
| License | MIT | MIT | Apache-2.0 | Apache-2.0 |
| GitHub stars | 20,182 | 58,728 | 42,233 | 31,923 |
| Releases in 90 days | 65 | 23 | 29 | 6 |
| Multiple agents working together | yes | yes | — | yes |
| Custom tools and functions | yes | yes | yes | yes |
| MCP support | yes | yes | yes | yes |
| A2A protocol | — | yes | yes | yes |
| Agent memory | yes | yes | yes | yes |
| Knowledge base (RAG) | — | yes | yes | yes |
| Sandboxed code execution | — | — | — | yes |
| Human confirmation | yes | yes | yes | — |
| Graphs and state persistence | yes | yes | — | — |
| Tracing and debugging | — | — | yes | — |
| Agent quality evaluation | yes | — | — | — |
| Voice and speech | yes | — | — | yes |
| Own server and Docker | — | — | yes | — |
| Images and documents | yes | — | — | yes |
| Own model: Ollama | yes | yes | yes | yes |
| Own model: LM Studio | — | yes | — | — |
| Own model: OpenAI-compatible API | — | — | yes | — |
Verified against project READMEs and documentation as of 09/26/2026: "yes" means the capability is stated there, a dash means it is not mentioned (which does not mean the capability is absent). Stars and releases — according to GitHub data.
Updating is worthwhile for those who work with realtime voice and have been waiting for GPT-Live or more reliable sessions, and worth a closer look for those who need typing throughout the agent's path and built-in quality evaluation. You can follow new releases of platforms for your agents in the CPU3D news feed.
