Vercel AI Gateway: what it is and when it earns its place

What Vercel AI Gateway actually is

It is a proxy with a model catalogue and a wallet. You send a request naming a model as provider/model, the gateway forwards it to whichever upstream can serve it, and the tokens are billed against a credit balance you top up in the Vercel dashboard. If the first upstream errors, it retries against another provider that serves the same model.

The part people get wrong is the hosting assumption. AI Gateway is a Vercel product, not a Vercel deployment feature. A Python script on your laptop, a Fly.io container, or a Kubernetes pod can all point at it. The only thing you need is a Vercel team account and an API key.

How it relates to the AI SDK

The AI SDK is Vercel's TypeScript library for calling models. AI Gateway is the default provider inside it, which is why AI SDK code written in 2026 passes a bare model string rather than importing a provider package.

import { generateText } from 'ai';

const { text } = await generateText({
  model: 'anthropic/claude-opus-5',
  prompt: 'Summarise this changelog.',
});

// swapping providers is a string edit
// model: 'openai/gpt-5.6-sol'
// model: 'google/gemini-3.1-pro'

You can use one without the other in both directions. The AI SDK can be pointed at a provider directly with that provider's own key, and the gateway can be called from Python, Go, or curl with no AI SDK anywhere. Treating them as one thing is the most common source of confusion in the docs.

curl https://ai-gateway.vercel.sh/v1/chat/completions \
  -H "Authorization: Bearer $AI_GATEWAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-opus-5",
    "messages": [{"role": "user", "content": "hello"}]
  }'

BYOK versus Vercel-billed credits

Two billing rails run through the same endpoint. On the default rail, Vercel authenticates with its own provider credentials and deducts the token cost from your AI Gateway Credits balance. On BYOK, you store your own provider key and Vercel authenticates with that instead, so the tokens land on your own provider invoice.

Both are advertised as zero markup, and as of August 2026 both are. The difference is entirely in the failure modes and the spend controls, and those are where BYOK gets uncomfortable.

providerOptions: {
  gateway: {
    byok: {
      anthropic: [{ apiKey: process.env.ANTHROPIC_API_KEY }],
    },
  },
}

Request-scoped keys bypass the dashboard-stored credentials entirely for that call, which is the right shape if you are a platform billing your own customers for their own model usage. The system-credential fallback still applies.

Coding agents through the gateway

This is the part of AI Gateway most relevant if you are here from a coding-agent search. Vercel ships per-agent compatibility endpoints and a CLI command that writes the config for every agent it finds installed.

vercel ai-gateway coding-agents setup

# or one at a time
vercel ai-gateway coding-agents setup --agent claude-code

export ANTHROPIC_BASE_URL="https://ai-gateway.vercel.sh/claude-code"
export ANTHROPIC_API_KEY=""
export ANTHROPIC_AUTH_TOKEN="your-ai-gateway-api-key"
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

Observability, and the features that cost extra

Spend by model, by project, and by API key comes with the product. Request traces and latency metrics live in the Vercel Observability tab. Budgets can cap a team, a project, or a single API key, with the BYOK exception noted above.

Three capabilities are off by default and bill on their own meters when you turn them on. They are cheap per unit and easy to leave running.

The per-request versions of the allowlist and ZDR are free. If you only need to restrict providers on some calls, use the only parameter in providerOptions rather than the team-wide setting, and the $0.10 per 1,000 disappears.

When it is the right call, and when it is not

Reach for it when

• You are already a Vercel team and want one invoice instead of five provider accounts.
• You are writing a Next.js or AI SDK app and want to A/B two models by editing a string.
• You need one endpoint that speaks Anthropic Messages and OpenAI Responses, because your agents are split across both.
• You want provider failover without writing the retry loop yourself.

Look elsewhere when

• Your agents run on your own machines and you already pay a subscription. Claude Pro, Claude Max, and ChatGPT plans are flat-rate; the gateway converts that into metered tokens.
• You need a hard cap on your own provider keys. BYOK spend cannot be budgeted, which rules the gateway out as a spend-control layer for BYOK teams.
• You need self-hosting. There is no on-prem build. LiteLLM or Bifrost is the answer to that question.
• You want an audit trail your compliance team owns. Trace Drains export OpenTelemetry, but the primary store is Vercel's.

Questions people ask

Keep reading

Continue through LLM gateways and routing with Vercel AI Gateway pricing: what zero markup does and does not cover.