Most teams reach for an LLM gateway the day a provider has an outage in production, or the day finance asks which feature is burning the token budget. Vercel AI Gateway is built around exactly those two moments — reliability and cost visibility — and it's the newest serious entrant in the space, so here's the quick, honest read.
What it is. A hosted gateway: one API key, one endpoint, and access to hundreds of models across 40-plus provider organizations. You stop integrating a separate SDK per provider and start swapping models by changing a string. It's native to the Vercel AI SDK and also speaks an OpenAI-compatible endpoint, so non-Vercel stacks can use it too.
Why it's worth a look. Two features usually cost extra elsewhere, and here they're the default:
- Automatic failover. Since April 2026 the gateway retries transient upstream errors — up to five attempts with configurable backoff — with no client-side change. Vercel reports this quietly rescues about 3.5% of requests and 5.1% of tokens, on the order of a trillion tokens a month that would otherwise have errored. That's uptime you didn't have to code.
- Built-in observability. A dashboard shows requests, token usage, and spend across every provider you route to, tied to your Vercel projects — so you can read cost per feature or per deployment without standing up separate monitoring.
What it costs. This is the headline: provider token rates pass through at 0% markup. The only base charges are optional team-wide controls — zero-data-retention and provider allowlists — at $0.10 per 1,000 successful requests. Compared with gateways that take a percentage of spend, a high-volume app can save real money here; Vercel makes its money up-selling adjacent infrastructure, not marking up your tokens.
How you start. Create an AI Gateway key in the dashboard, keep your AI SDK code, and set the model to a provider/model slug:
import { generateText } from "ai";
const { text } = await generateText({
model: "anthropic/claude-sonnet", // swap to "openai/gpt-4o" with one edit
prompt: "Summarize this support ticket in two sentences.",
});The request now routes through the gateway, failover and all, and shows up in the spend dashboard immediately. Not on the AI SDK? Point any OpenAI-compatible client at the gateway's base URL instead.
Who it's for — and who it isn't. It's the strongest pick if you're already on Vercel or the AI SDK and you want reliability and spend visibility with essentially zero ops. If your priority is the widest possible model marketplace under one credit pool, OpenRouter's catalog is broader; if you need every request to stay inside your own infrastructure with hard per-team budgets, the self-hosted, MIT-licensed LiteLLM is the tool to run. For the fuller field of hosted-vs-self-hosted trade-offs, see OpenRouter vs LiteLLM vs Cloudflare AI Gateway.
The one-line version: Vercel AI Gateway makes failover and cost transparency the default instead of the upgrade, and charges nothing to pass your tokens through. If reliability and a clean spend dashboard are what pushed you toward a gateway in the first place, it's the lowest-friction way to get both.
