AI gateway on the Vercel platform
INFRO vs Vercel AI Gateway
Vercel AI Gateway is one endpoint across a large model catalog, built into the Vercel platform and the AI SDK. It takes no markup on tokens, sells model access through its own AI Gateway Credits with a free tier on a subset of models, and adds budgets by team, project, key and member, observability, zero-data-retention routing and regional inference. For a team that deploys on Vercel and writes against the AI SDK, it is the shortest path from nothing to a working model call, and that is a real advantage rather than a convenience.
- You deploy on Vercel and build with the AI SDK. A project's deployments authenticate with their own OIDC tokens, so there is no gateway key to create, store or rotate, and project budgets attach to those deployments directly.
- You want model access without holding provider accounts at all. Vercel buys the inference and bills it against its credits; INFRO only routes to provider accounts you already own, so it has nothing to offer a team that has none.
- You want to start at zero. Vercel's free tier covers a subset of models; every INFRO plan is paid, with a 14-day trial.
- Per-member budgets with defaults are what you need. Vercel's default budgets give every project, key or member its own allowance the moment it exists, including ones created later, with email alerts at 50, 75 and 100 percent.
- You need it working today. AI Gateway is generally available and self-serve; INFRO is in private early access.
This section is not a courtesy. If one of these describes you, pick Vercel AI Gateway — a customer who adopts the wrong layer on our advice is worse for us than one we never had.
Where the two actually diverge.
- Whose account carries the request
- Vercel offers both arrangements: requests on Vercel's own provider accounts, paid from AI Gateway Credits, or on your own keys with no markup. Its documentation says bringing your own key is a paid-tier feature that needs purchased credits, and that when a request on your credentials fails it is retried on Vercel's system credentials and charged to that balance — with zero-data-retention on, only to system credentials whose providers have ZDR agreements. That fallback is a reasonable design for availability. It also means a request you meant to run under your own contract can complete under someone else's. INFRO holds no provider accounts and no balance, so there is nothing to fall back to: failover moves between the connections you made, in the order your route sets, and a request no connection can serve fails rather than leaving your accounts.
- What a budget counts
- Both products refuse requests with a 402 once a dollar limit is used up, at similar scopes. The difference is which spend the limit sees. Vercel's budgets and BYOK pages, read on 2026-10-13, say spend through your own provider keys is metered separately and not counted in any budget, so a budget cannot cap it; the budgets page also describes each budget as a soft cap, checked at the start of a request, so the request that crosses the limit still completes. In INFRO every request runs on your own keys, so every request is what budgets govern: organization, project, member and key, on every plan. A blocking budget admits a request only if its worst-case estimated cost fits the room left, counting requests still in flight, and a model with no known price is refused under a blocking budget unless the owner allows unpriced traffic. All of it counts INFRO's estimate of provider cost, not your provider's invoice.
- Estimates at your rates, then the provider's own number
- A gateway that sells inference prices requests at its own rates; one that only routes your accounts has to price them at yours. INFRO estimates each request from the vendor's published rate or, from the Team plan, from the rates your contracts carry, stored per model and optionally per provider. On the Business plan it compares the month's estimate with OpenAI's and Anthropic's own cost APIs and states the variance.
- Part of a platform, or a product of its own
- Vercel's gateway is one surface of a deployment platform, and much of what makes it good — OIDC authentication from deployments, the AI SDK, one Vercel bill — comes from that. INFRO is accountable to nothing except operating AI across whoever runs it, wherever their applications are deployed and whichever providers they hold contracts with. If your applications and your model access all live on Vercel, that independence buys you little. If they do not, it is the point.
Side by side.
Our column carries dashes too. A comparison where one side is all checks is one nobody believes, and rightly.
| Capability | Vercel AI Gateway | INFRO |
|---|---|---|
| One API across many providers | yeslarge catalog; Vercel's accounts or yours | yeswhatever your connected providers serve |
| Model access without your own provider accounts | yesAI Gateway Credits; free tier on a subset | noyour own provider accounts only |
| Your own provider keys | partialpaid tier; purchased credits required | yesthe only mode; connected per environment |
| Failover stays on your own accounts | nodocumented retry on Vercel's credentials | yesacross your own connections only |
| Budgets that refuse a request | partialteam, project, key, member; documented as a soft cap | yesorganization, project, member, key; 402 |
| Budgets that cover spend on your own keys | noBYOK spend not counted, per their docs | yesevery request is on your own keys |
| Zero-data-retention routing | yesper request or team-wide | noyour own provider agreements apply |
| Authentication from your deployments | yesOIDC tokens on Vercel | nogateway keys |
| Free tier | yesa subset of models | nopaid plans, with a trial |
| Available today | yes | noprivate early access |
Our reading of each product's public documentation as of October 2026. We do not quote competitors' prices — theirs change and a wrong number about someone else's business is worse than none. Think a row is wrong? Tell us and we will fix it.
Work out what the difference is worth
Estimate what your current model mix costs at your providers' list rates and the plan it needs, or have us read your real usage and tell you where we would and would not help.