Baseten vs AutoGPT

The catalog side of Baseten churns constantly — Kimi K3, Inkling, GLM 5.2 Fast added, with GLM 5.1, GLM 5, Kimi K2.5, and Nemotron Super 120B deprecated in the same month. Underneath that rotation, two more durable things landed: a Fast tier that serves identical weights on dedicated capacity for higher sustained per-user throughput, and Baseten for Model Labs, which sells serving and distribution infrastructure to the labs producing the models. The control plane filled out alongside — org-scoped key management, admin visibility into personal keys, GPU usage accounting, and programmatic logs, metrics, and audit logs.

Baseten is climbing from the bottom of the stack toward the middle. The Fast tier says the differentiator is no longer which models are available but how they're served under agentic load, where sustained per-user throughput matters more than headline latency. Model Labs goes further and inverts the customer: instead of application teams renting inference, the labs themselves rent distribution. Together they move Baseten from interchangeable GPU capacity toward being the layer a model reaches the market through.

The Fast tier arriving with a single model makes extending it across the catalog the obvious next step, and Model Labs suggests lab-hosted models appearing in Model APIs under their originators' branding. The deprecation cadence indicates the catalog will keep rotating quickly rather than accumulating.