Baseten joins OpenAI B2B Marketplace as open-model inference provider

What the partnership offers

Under the partnership, OpenAI enterprise customers can run open models hosted by Baseten without setting up a separate spending arrangement. Baseten says they can apply their existing OpenAI commitment to these models, either inside Codex or via the Responses API.

Baseten frames this as a way to deliver "even more intelligence per dollar" through that existing commitment. Access to open models natively in Codex is not yet open to everyone: the company asks interested customers to join a waiting list.

The company argues that organizations can now optimize agentic workflows across open and closed models and route each task to the best-fit model. It points to OpenAI's Codex and GPT models as two of the most popular choices for scaling code-generation workflows.

Why Baseten says multi-model matters

Baseten describes agentic coding as the first multi-model workflow at global scale. According to the company, leading AI adopters are now using a mix of models at Pareto frontiers across intelligence, cost, latency and capacity. Top engineering teams, it says, send tasks through static or adaptive routers to whichever model gives the best balance for the job.

Baseten also notes that the cycle between new open and closed models has shrunk to weeks, so no single model stays the best tool for every job for long. Companies therefore need instant access to the latest models, with fast and reliable inference that fits into their harnesses, gateways and other tooling.

The infrastructure behind it

Baseten lists several capabilities aimed at code generation, which it says is demanding at the inference layer because of multiple turns, long prompts over large repositories and huge context windows.

• Performance: optimization work that spans everything from tooling down to the engine level.
• Capacity and reliability: Baseten says it runs on more than 90 clusters across 20+ clouds, with active-active deployments so that losing a provider or region reroutes traffic instead of stopping work. It cites relationships with 200+ compute providers.
• Day zero model access: major open models launch on Model APIs the day they ship, and trying a new one is a single-line code change, according to the company.
• Developer and agent experience: Blaxel, now part of Baseten, gives every agent its own sandbox to run the code it writes.
• Enterprise governance: inference runs on US-based infrastructure with zero data retention (ZDR) for all prompts. Deployments can be pinned to specific regions, with fine-grained AuthN and AuthZ and usage views across any model, user or key.
• Research and engineering support: forward-deployed engineers tune deployments to latency targets, and applied researchers run post-training with customers. Baseten says LangChain trains custom models for LangSmith Engine this way with Baseten Loops.

What it means for enterprise customers

For organizations already committed to OpenAI, the change means open models can sit alongside GPT models in the same procurement and tooling path. Baseten says this lets teams mix open and closed models in coding workflows while drawing on one commitment.

The company adds that it will keep building toward access to open frontier models on day zero, fast and reliable inference, and guaranteed capacity in its cloud or the customer's own. Customers who want native open-model access in Codex can join the list, and others can contact a Baseten engineer.