OpenAI Agents SDK adds Codex as an opt-in provider
The Python Agents SDK's first 1.0.0 alpha lets developers pick Codex per run through a new model provider, with wheels distributed only through GitHub Releases.
What shipped
On 20 November, OpenAI published Agents SDK v1.0.0-alpha.1, a pre-release of the Python SDK in which the OpenAI Agents SDK Codex provider becomes an opt-in choice for a single run rather than a fixed setting. The notes describe it as the first preview from the feat/simple-codex branch. The prebuilt openai-agents wheels, numbered 1.0.0a1, are available only through GitHub Releases, and the details sit on the v1.0.0-alpha.1 release page.
This is not a new model. It is a second way to hand a model to the SDK, chosen per run rather than fixed for the whole app.
Choosing the provider per run
Import CodexProvider from agents.models.codex and create an instance. Pass it through RunConfig(model_provider=provider) to the Runner. By default Codex uses the saved login on the machine; CodexProvider(api_key=...) switches to API key authentication. The provider selects the model for agents that name one or fall back to the default model, and the notes say the same agent can run with either provider, which makes model choice a configuration decision.
Under the hood the wheel embeds unmodified Codex Core rust-v0.166.0 through PyO3. Calling Core directly removes codex app-server startup and JSON-RPC from the path. One provider instance shares a ThreadManager across calls, each call gets a fresh thread, history comes from an SDK session, and you close the provider with await provider.aclose().
Streaming text, handing off to a specialist, pausing for approval before a Python tool runs and validating typed output are all listed as working. Native commands follow Codex's sandbox and approval settings, and a native approval prompt ends the model call.
Install and platforms
Use a virtual environment with CPython 3.10 through 3.14. Wheels cover Linux x86-64 and Linux ARM64 (glibc 2.28 or newer), macOS 11 or newer on Apple silicon, and Windows x64. Each download link carries the SHA-256 digest pip checks, and checksums for all four wheels are published. The install line is python -m pip install --only-binary=:all: with the wheel URL.
For the wildlife example, check out the tag and run the script from the repository root with the wheel installed. If the only credential available is a saved Codex login, set OPENAI_AGENTS_DISABLE_TRACING=1 in the shell first.
OpenAI reports that the SDK suite passed on the merged branch, that CI built all four wheels and tested fresh installs on CPython 3.10 and 3.14 on each platform, and that all four downloads were verified against SHA256SUMS. A live run of the wildlife example used the Linux x86-64 wheel on CPython 3.12: it identified an American bison, called the Python catalog tool and returned a typed report with record ID synthetic-bison-001.
Fixes and the migration path
Automatic session compaction now uses local conversation history for Codex, including after an approval pause, and needs its own OpenAI client or credentials. Saved sandbox manifests keep inline binary files when a run resumes; if you share manifests, upgrade every reader before saving non-UTF-8 content, and use LocalFile when older SDKs have to read them.
Anyone on the v0.24.0-alpha.8 harness preview has work to do. Agent.run() and Session.submit() move to Runner.run() or Runner.run_streamed(), with CodexProvider set in the run config. Code using agents.classic should import Agent, Runner and RunConfig from agents. Pending native runs and approvals should be finished inside the alpha.8 environment, because its SessionState checkpoints differ in format from the Runner's RunState, and history should start in a fresh SDK session.
Limits and open questions
This is a pre-release with no stated stability guarantee and no PyPI install path in the notes. Nothing is said about Codex usage pricing, rate limits, or whether the provider works with models other than Codex. There is no timeline for a stable 1.0.0, and the notes list only the four wheels above, none for Windows on ARM. The instruction to finish approvals before switching suggests the checkpoint change is a one-way break within the alpha series, though the source does not say so outright.
What this suggests for app builders
Read as a signal rather than a product, the API shape suggests model choice in agent frameworks is moving toward a per-run setting, so one agent definition can be kept while the engine behind it changes. That is inference, not a claim from OpenAI. The practical takeaway is to keep model configuration out of business logic so a provider switch stays a config change. Draftship's notes on adding AI to your app cover that same separation; Draftship is not affiliated with OpenAI and does not support anything in this release.
Primary sources
- Agents SDK v1.0.0-alpha.1 · OpenAI Agents SDK releases, Nov 20, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.