Baseten launches model router, sandboxes, and training
Baseten is adding a model router, agent sandboxes and generally available serverless training to its inference platform, a bid to let enterprises run their AI stack through one vendor instead of several.
Baseten announced three launches this week alongside a website redesign. The company says teams that piece their AI systems together from separate vendors lose time on integration, add latency and suffer downtime when one part fails.
Grid, a gateway, router and guardrail product, enters public preview the day after the announcement. It is aimed at enterprise spending on coding agents. The gateway puts model usage and spend in one place, the router sends each task to a suitable frontier or open-weight model, and guardrails set company-wide rules on access. Baseten says Grid works with any harness, including Claude Code, Codex, OpenCode and Pi, and that its routing tries to maximize KV cache hits in multi-turn agent sessions. Opendoor already routes every AI request across the company through Grid, according to Baseten. Fuller details are due at launch.
Sandboxes move from private to public preview on Thursday. They build on Baseten's recent acquisition of Blaxel and are described as stateful, pause/resumable and VM-isolated, able to scale to thousands of instances in seconds. They will run alongside Baseten's inference. Terms of the Blaxel deal were not disclosed.
Loops, Baseten's serverless SFT and RL service for open-weight models, becomes generally available on Friday after a public preview. Its SDK is compatible with Tinker. Customers can run it on single-tenant capacity, serve trained models through Baseten's Dedicated Inference and keep ownership of their weights. Baseten names Datadog, Harvey, Lovable and Parallel among the teams it helps with post-training.
The company did not give pricing for any of the three products.