Baseten launched Serverless Loops training

Baseten launched a new Serverless Loops training product with per-token pricing for GLM-5.3, GLM-5.3-Flash and Qwen3.6 35B-A3B across prefill, cached, sample and train rates, each carrying a 20% limited-time introductory discount, per the pricing page Pulse saved on Nov 7, 2026. Listed rates per 1M tokens are $3.02 prefill, $0.60 cached prefill, $9.72 sample and $11.66 train for GLM-5.3; $0.67, $0.13, $1.69 and $4.46 for GLM-5.3-Flash; and $0.36, $0.07, $1.07 and $1.71 for Qwen3.6 35B-A3B, per Baseten's pricing page on Nov 7, 2026. The product sits alongside Baseten's existing Model API and Sandboxes usage pricing as a new, separately billed line.

Serverless Loops puts Baseten in the business of selling model training by the token, not just inference — a new line alongside the Model API and Sandboxes usage pricing it already runs. For a product leader bundling AI capability into their own product, this is a signal that infrastructure providers are pricing fine-tuning and continued-training workloads the same way they price inference: usage-metered, per-model, with an introductory discount to pull in early volume.

Watch how long the limited-time intro discount lasts — Baseten's page gives no end date — and whether Baseten extends per-token training pricing to more of its models; that would confirm this is a platform-wide pricing primitive, not a three-model launch.