Route AI model requests through Vercel AI Gateway with provider fallback ordering and BYOK, then observe usage and spend

Documented steps

1. Create an API key at the AI Gateway API Keys dashboard page and export it as AI_GATEWAY_API_KEY (the AI SDK reads this automatically); alternatively, on a Vercel deployment use the auto-provided VERCEL_OIDC_TOKEN as a fallback auth method with no key management.
2. Call a model using the unified model-string format 'provider/model' with the AI SDK's generateText/streamText — a plain string model id automatically routes through the AI Gateway provider. Check the gateway's model list page for currently available provider/model slugs.
3. To control provider fallback order, pass providerOptions.gateway.order: ['bedrock','anthropic'] (try in this order) and/or only: ['bedrock','anthropic'] (restrict to only these providers); use sort: 'cost'|'ttft'|'tps' to auto-rank by cost, latency, or throughput instead of a fixed order.
4. AI Gateway automatically retries a failed request against other providers/system credentials for reliability, even without explicit order/only config, by default choosing providers based on recent uptime/latency.
5. For BYOK, add provider credentials at Team > AI Gateway > Bring Your Own Key (dashboard) — Vercel tries your key first and falls back to system credentials (billed to your credits) on failure; or pass request-scoped credentials via providerOptions.gateway.byok: { anthropic: [{ apiKey: process.env.ANTHROPIC_API_KEY }] } (credential shape varies per provider, e.g. Bedrock needs accessKeyId/secretAccessKey, Vertex needs project/location/googleCredentials).
6. BYOK requires the team to have purchased AI Gateway credits (paid tier feature); fallback-to-system usage after a BYOK failure is billed against that credits balance.
7. Monitor usage in the dashboard's AI Gateway Overview: Usage tab shows Requests by Model, Time to First Token, Input/Output token counts, and Spend, scoped at team or project level; Requests tab breaks down by project and API key with request count, avg tokens, P75 duration/TTFT, and cost.
8. For per-request debugging, open the Logs page under AI Gateway to search by request ID, filter by model/provider/status code, follow requests live, inspect routing, and export filtered results as CSV/JSON.
9. Official docs verified: https://vercel.com/docs/ai-gateway/authentication-and-byok | https://vercel.com/docs/ai-gateway/models-and-providers/provider-options | https://vercel.com/docs/ai-gateway/observability-and-spend/observability | Reference: https://vercel.com/docs/ai-gateway