Framework Update: Pydantic AI v2.46.0

TL;DRThree things worth knowing about Pydantic AI v2.46.0: (1) the previous release (2.45.0) introduced TypeSafeModel — a provider for TypeSafe's Jev, a classifier that answers typed questions instead of writing text — and this release fills in what it couldn't do yet: filling tool call arguments and picking a type before filling a union output; (2) a new `typesafe_boolean_threshold` turns the yes/no decision boundary from a fixed distance-from-0.5 into a tunable parameter; (3) `supports_text_output` lets `LLMJudge` and `GEval` run on models that don't produce text at all, so Jev can now serve as the judge model in evals. No breaking changes.

#ai-agent #framework #daily #pydantic-ai

Series: AI Framework Changelog (25 / 25)

Framework Update: Microsoft Agent Framework python-1.19.0

🌏 中文版

Release Info

Item	Value
Framework	Pydantic AI
Version	v2.46.0
Previous version	v2.45.0
Release date	2026-09-19
Release Notes	GitHub Release
GitHub	pydantic/pydantic-ai
Stars	20k

Why this release matters

The previous release (2.45.0, 2026-09-17) merged TypeSafeModel, a provider that connects to TypeSafe's Jev classifier. Jev isn't a language model — it doesn't write text. You give it a passage and a set of typed questions (each field of a Pydantic model becomes one question), and it answers every field directly with a confidence score. Because it skips sequential token generation, an AlphaSignal ticket-triage benchmark measured a 227ms median latency for Jev against 1,415ms for gpt-5.6-luna — roughly 6x faster. But 2.45.0's limits were just as clear: str outputs, native file inputs, and tool calls that require generated arguments were all unsupported, raising a UserError on contact. 2.46.0 closes exactly two of those gaps — filling tool arguments and handling union output types — while also making Jev's confidence threshold tunable and, notably, usable as an eval judge. For teams already routing decisions through TypeSafeModel, this isn't a standalone feature drop; it's the previous release's gaps getting filled in, one version later.

Key changes

TypeSafeModel fills tool arguments (when Jev can express them): Jev previously only filled output_type fields. Starting in 2.46.0, when a tool's argument schema falls within Jev's supported question types (bool, Literal, Enum, a float bounded to [0, 1], and similar), tool-call arguments can be filled by Jev too, without switching to a text-generating model mid-run
TypeSafeModel fills a union output type by picking the type first: when output_type is a union, Jev first answers "which type applies," then fills that type's fields — union types were previously out of scope for Jev
typesafe_boolean_threshold: yes/no decisions used to be based on a fixed distance from 0.5; this release makes that threshold configurable, useful when the cost of false positives and false negatives is asymmetric — you can tighten or loosen the decision line directly instead of wrapping a post-processing layer around the agent
supports_text_output on ModelProfile lets LLMJudge/GEval run on models with no text output: Pydantic AI's built-in eval tools, LLMJudge and GEval, previously assumed the judge model could write out its reasoning as text. This release lets them recognize a model that can't, so Jev — a pure classifier — can now serve as the judge, scoring agent outputs with a much faster classifier instead of a generative one
Choices helper: builds a set of described options at runtime, for dynamically generated classification/selection fields
Enum options described via member docstrings (UseEnumMemberDocstrings): each enum option can carry its own docstring as its description, which matters for classifiers like Jev that need a clear question description per field
RealtimeSession.wait_for_playback(): real-time voice sessions get a method to wait for playback to finish; the official docs examples were updated to wait for the reply to finish before closing
TemporalDurability adds event_stream_topic: streams agent events out through Temporal's Workflow Streams, contributed by the community — one of the few changes in this release unrelated to TypeSafeModel

Breaking Changes

No breaking changes in this release.

Migration Guide

Upgrade directly, no code changes required:

pip install --upgrade pydantic-ai==2.46.0

Code changes are only needed if you want to use the newly filled-in capabilities — for example, letting Jev fill tool arguments:

from pydantic_ai import Agent
from pydantic_ai.models.typesafe import TypeSafeModel

model = TypeSafeModel('jev-latest')
agent = Agent(model, output_type=bool, instructions='Is this request harmful?')

# Starting in 2.46.0, tool-call arguments that fall within Jev's
# supported question types (bool / Literal / Enum / a 0-1 float, etc.)
# can be filled by Jev too, without switching models mid-run

Comparison with other frameworks

LangGraph and CrewAI are largely about orchestrating multiple text-generating models. Pydantic AI's TypeSafeModel opens a different path: swap out the steps in an agent's decision chain that are really just classification — continue or not, which branch, a risk score — for a non-generative discriminative model, while keeping that step inside Pydantic's type validation and provider interface. By wiring tool arguments and union types into TypeSafeModel in 2.46.0, Pydantic AI signals this isn't a one-off demo integration but an intent to let classifiers and generative models coexist in the same agent, switched in as needed.

Today's takeaway

I used to assume "model" in an agent framework defaults to "something that generates text." Watching Jev — a classifier that never writes text, only answers typed questions — get wired in as a first-class citizen of Agent, sharing the same provider interface and type validation as any language model, changed that: a lot of the steps in an agent's decision chain are classification problems, not generation problems, and forcing a generative model onto them just means paying token-generation latency and cost for nothing.