Vercel AI Updates: September 22, 2026

1. Jev Is Now Reachable From AI Gateway Without the AI SDK

Vercel opened two more paths to Jev through AI Gateway: an existing TypeSafe client can be pointed at the Gateway without changing its evaluation calls, and there is now a plain HTTP API callable from any language. The AI SDK route remains for TypeScript applications. Dropping the SDK requirement is what makes Jev usable from Python and Go services that had no way in. Source

2. Xiaomi’s MiMo V2.6 Family Arrived on AI Gateway

Vercel added MiMo V2.6 Pro, MiMo V2.6 Flash, and MiMo V2.6 Pro UltraSpeed from Xiaomi to AI Gateway. The family combines coding, reasoning, and tool use with native text, image, audio, and video understanding, carries a 1M token context, and supports structured outputs with up to 128K output tokens. The context size is aimed at long repositories, tool traces, and multi-session agent work. Source

3. Grok 4.7 Landed on AI Gateway at 40 Percent Off Through September 27

Vercel made Grok 4.7 available across AI Gateway, fx, and eve, with the discount applied automatically on calls to spacexai/grok-4.7 until September 27. The model has a 500K token context window and exposes low, medium, high, and xhigh reasoning levels. The reasoning-level control is the part worth wiring into an agent loop, since it trades latency against depth per call. Source

4. Vercel Connect Added a Managed Microsoft Teams Connector

Vercel shipped a managed Teams connector that gives an organization a bot its apps and agents run. People can mention it in channels or message it directly, and application code receives the message and replies as the bot. Vercel registers the Entra app and Azure Bot resource as part of the managed connector, which removes the usual Azure setup from the path. Source

5. Deployments Now Report Billable Duration and CPU Minutes

Vercel surfaced billable duration and CPU minutes per deployment in the dashboard, vc inspect, and the REST API. Billable duration is build plus post-build time rounded up to the next whole minute, and CPU minutes are that figure multiplied by the machine’s vCPUs. Per-deployment attribution makes it possible to find which builds are driving spend instead of reading a monthly aggregate. Source