On November 16, 2026, Vercel announced in its changelog that AI Gateway now supports Private Inference (paired with documentation marked last updated September 16, 2026), allowing teams to connect their own fine-tuned, self-hosted, or prerelease model endpoints directly to AI Gateway.

How Private Inference Works & Protocol Compatibility

Connecting Custom Endpoints (Distinct from BYOK): Unlike BYOK—which supplies your own keys for providers already supported by AI Gateway—Private Inference registers custom endpoints and models unknown to the public catalog. Private models remain scoped to your Vercel team, never appear in the public catalog or /v1/models, and share unified auth, SDK/API interfaces, and observability (configurable via Vercel CLI 65.16.0 or later and usable in connected coding agents such as Claude Code, Codex, and Cursor).

Invocation Prefix & Supported Formats: Models are invoked with a private/ slug prefix and support OpenAI Chat Completions, OpenAI Responses, Open Responses, and Anthropic Messages formats.

Quotas, Pricing & Team Permissions

Endpoint Limits, Credits & Fallback Pricing: Endpoints must be publicly reachable over HTTPS, capped at 25 private providers and 100 private models per team, and require a positive balance of purchased AI Gateway Credits. By default, requests naming only private models with no Vercel-provided catalog fallback incur a list-price fee of $0.20 per 1,000 requests ($200 per 1,000,000 billable requests; custom enterprise rates are available via sales), whereas chains with a catalog fallback and Fallback to Vercel Keys enabled incur no fee. Creating, editing, deleting, or enabling Zero Data Retention requires Member or Owner permissions.