Pydantic AI 2.33.0 requires Anthropic SDK 1.0 / httpx2

If Anthropic models in Pydantic AI stopped working in the last day, you are not imagining it. Version 2.33.0 is a compatibility firefight, not a feature drop. It exists because anthropic 1.0.0 reached PyPI and every earlier Pydantic AI release could install that SDK without supporting it.

On August 20, 2026, Anthropic’s Python SDK 1.0.0 landed on PyPI. The v2.33.0 GitHub release notes describe that package as rebuilt on httpx2, with legacy httpx support removed. Every earlier Pydantic AI release allowed 1.0.0 without supporting it. The notes call out v2.32.2 by name: it was cut a few hours before 2.33.0, and it still permitted the new SDK. A fresh or unpinned install of pydantic-ai[anthropic] could resolve anthropic 1.0.0 and then fail at runtime when using an Anthropic model.

Pydantic tagged v2.33.0 on 2026-08-20 and published the GitHub release on 2026-08-21 (04:53:40Z). The same morning, @pydantic posted that Pydantic AI version 2.33.0 is out, linking that tag. Versus v2.32.2, the changelog is essentially pull request 7657: “Use httpx2 for Anthropic clients,” by @dsfaccini.

There is no feature tour in this tag. If you came looking for new agent capabilities, this is not that drop. It closes a breakage window. The maintainers apologize for that window in the release notes.

The breakage is a resolver hole

The problem is not a quiet API tweak inside Pydantic AI. Through 2.32.2, the project’s Anthropic extra could accept anthropic 1.0.0 even though the library did not support that SDK. Installers that take the newest matching version could therefore pull 1.0.0. The install could look fine. Failure showed up later, when code actually used an Anthropic model.

That is why this presented as a last-day outage rather than a failed install. Environments created or refreshed on August 20 without an Anthropic pin were first in line. Environments that never re-resolved may still be on a 0.x SDK until the next install. The release notes are explicit that “every earlier” pydantic-ai release had this hole, including 2.32.2 hours before the fix.

v2.32.2 is not a safe harbor. Being on “the previous release” does not protect an unpinned pydantic-ai[anthropic] extra. If Anthropic is on your path and you resolved extras around August 20, treat 2.33.0 as the supported cut.

Two exits

The notes give two supported responses. Pick one.

Upgrade. Move to pydantic-ai v2.33.0. That release requires and supports anthropic>=1.0.0. This is the intended fix if you can take Anthropic’s 1.x SDK and the httpx2 stack it ships on.

Stay put. If you need to remain on an older Pydantic AI, pin anthropic<1. That keeps you off 1.0.0 and on a legacy-httpx SDK the older library can still use.

Do not assume an unpinned pydantic-ai[anthropic] extra will do the right thing on a pre-2.33.0 line. That extra is exactly what the notes say could resolve 1.0.0 and then fail at runtime. Mixing an old Pydantic AI with Anthropic 1.0.0 is the broken state. Mixing 2.33.0 with a forced anthropic<1 fights the new requirement. Either move the library forward, or hold the SDK back.

Custom `http_client` is a hard type change

Default construction is not the only break. If you pass your own http_client to AnthropicProvider, it must now be an httpx2.AsyncClient. The 1.x Anthropic SDK rejects legacy httpx clients at construction.

Rejection happens when the provider is constructed, not on the first model call. If the object you pass is still a legacy httpx client, 2.33.0 will not accept it. You only need this swap if you inject a client. If you never set http_client, 2.33.0’s own Anthropic path is what PR 7657 changed. If you do inject one, the object has to be an httpx2.AsyncClient. That requirement is in the release notes in those words.

The changelog is one compatibility PR

Under What’s Changed, the release lists a compatibility note: use httpx2 for Anthropic clients, @dsfaccini, PR 7657. The full changelog is the compare from v2.32.2 to v2.33.0.

Plan around that delta only. Do not budget unrelated migrations into this tag. Do budget for the Anthropic extra, the SDK pin or bump, and any custom http_client you already pass to AnthropicProvider.

Who should act

Treat this as urgent if any of the following is true:

You install pydantic-ai[anthropic] without pinning anthropic.
You resolved dependencies on or after August 20, 2026, while still on v2.32.2 or earlier.
Failures appeared only when an Anthropic model ran, after an otherwise successful install.
You pass a custom http_client into AnthropicProvider.

If you do not use Anthropic models through Pydantic AI, the notes do not claim this cut affects you. The warning is scoped to Anthropic, httpx2, and AnthropicProvider.

Confirm the installed pair together: pydantic-ai and anthropic. On 2.33.0 you want anthropic 1.0.0 or newer. On anything older, you want anthropic below 1. A version check that only looks at pydantic-ai will miss the hole the notes describe.

How to read the official notes

The GitHub release is the source of the compatibility warning, the two-path guidance, the http_client type requirement, and the pointer to PR 7657. The @pydantic post on 2026-08-21 is the public announcement that 2.33.0 is out, and it links that tag.

They agree on the cut. 2.33.0 is the supported Anthropic 1.0 / httpx2 line. Earlier releases, including 2.32.2 hours earlier, could pull 1.0.0 without supporting it. The project’s own guidance is short: upgrade to 2.33.0, which requires and supports anthropic>=1.0.0, or pin anthropic<1 if you stay on older pydantic-ai. Custom clients must be httpx2.AsyncClient, because the 1.x SDK rejects legacy httpx at construction.

That is the operational contract. Apply it before the next environment rebuild.

Read the v2.33.0 GitHub release notes, then either upgrade pydantic-ai to 2.33.0 or pin anthropic<1 on the older line you are keeping. If you pass http_client to AnthropicProvider, change it to an httpx2.AsyncClient before you ship.