
Image: Vercel
Why it mattersTeams running a private endpoint beside a public catalog stopped juggling two sets of keys, client code and logs. Private and catalog models share one pipe.
A team running a fine-tuned model next to catalog models has kept two keys and two client libraries. Vercel says its AI Gateway now carries both through the same pipe. The company calls the new mode Private Inference. A team registers its endpoint as a private provider on the AI Gateway dashboard. Its models then get a private/<slug> ID. Any client that calls the AI Gateway can call them.
Sam Chitgopekar and Jerilyn Zheng announced the change on the Vercel changelog. The write-up names what the Gateway brings to a private endpoint. Chat Completions, Responses and Messages all work. Vercel says the Gateway translates each request to the endpoint's own format. So switching between a catalog model and a private one only changes the model ID.
What it covers
Private models stay scoped to the Vercel team. They do not show up in the public catalog. Vercel says credentials are stored encrypted. A team can turn on Zero Data Retention per provider.
The Gateway's observability extends to private calls. Vercel says requests to a private model appear in logs, trace drains and custom reporting alongside catalog traffic. Its charts break down request counts, tokens and time to first token by model. A team can set a fallback to a catalog model if the private endpoint fails on a request.
Where coding agents fit
Private Inference is wired into the coding-agent pipe that already runs through the AI Gateway. Vercel names Claude Code, Codex and Cursor, with "other coding agents connected to AI Gateway" in the same list. A developer using one of those agents against a fine-tune can point the agent at private/<slug>. The fine-tune comes back. This is the smallest practical payoff for a reader on a project.
The CLI has a one-line hook for setup. Vercel says version 65.16.0 or later accepts vercel ai-gateway private-inference providers create and a matching models create call. That cuts the "register a new endpoint" job to two shell commands.
What it does not cover
The post does not list every endpoint format Private Inference accepts, every model type, or the billing specifics. Vercel points to the Private Inference documentation for those. The feature is also only useful to teams already on AI Gateway. A team on another gateway product sees nothing from this change until it migrates.
The pattern is the interesting part. A platform that started as a router for catalog models is extending to the models a buyer brings with them. That reduces one of the real reasons a developer declined a managed gateway in the past.
Source
Vercel changelog: AI Gateway now supports Private Inference.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


