AI Gateway FAQ

Answers to common questions about AI Gateway, including request errors, pricing, compatibility, model availability, and data handling.

Match the response status, type, and message to its cause. A 403 has several possible causes, so check the error message before choosing a resolution.

No. AI Gateway charges the provider's list price with no platform markup on tokens, on both the free and paid tiers. The same zero-markup pricing applies to BYOK requests. See AI Gateway Pricing.

Token prices match provider list rates, and two mechanisms bring them below list. Some models are priced below provider list for every team, and volume commitments can negotiate custom discounts with invoice payment. A monthly free credit covers smaller workloads, and budgets cap spend per team, project, API key, or team member. Requests that hit a provider's prompt cache are billed at that provider's reduced cache-read rate, which the model catalog's per-endpoint pricing shows separately.

The AI Gateway Overview charts usage and spend by model, and Logs shows the cost of every request. Two REST endpoints cover the same ground without the dashboard. GET /v1/credits returns your credit balance and lifetime spend, and GET /v1/generation returns the cost of one request. For aggregates grouped by model, user, tag, provider, or credential type, use Custom Reporting, and set budgets with spend alerts to hear about usage before it grows.

No. The AI SDK is one path among several. You can keep an existing OpenAI or Anthropic SDK and change its base URL, call the OpenResponses HTTP API, send requests to the REST endpoints directly, use the AI SDK for Python, or go through a framework integration such as LangChain, LiteLLM, or Pydantic AI. You can also route coding agents and chat platforms through AI Gateway.

Yes. New models land in the catalog continuously, and you can verify freshness instead of taking a claim for it. Every model entry in GET /v1/models carries its provider release date in released:

curl -fsSL https://ai-gateway.vercel.sh/v1/models | jq '.data | sort_by(.released) | reverse | .[0:5][] | .id'

You can also browse the model list, which shows the same catalog with pricing and capabilities.

A slug that ends in -fast, such as anthropic/claude-opus-5-fast, is fast mode: the same model on a faster serving path at a higher per-token cost, with fallback to the base model when the fast path is unavailable. Using the fast slug is equivalent to setting the speed option on the base model.

Only the suffix carries that meaning. A provider's own model name can contain "fast" elsewhere without being a fast-mode variant, such as google/veo-3.1-fast-generate-001 or spacexai/grok-voice-think-fast-2.0. Those are separate models with their own pricing, which can be lower than the model they are named after rather than higher.

No. An AI Gateway API key authenticates requests from any environment, including local development, CI, and other cloud providers. Applications deployed on Vercel can authenticate with OIDC tokens instead of managing a key, but nothing about AI Gateway requires deploying to Vercel.

No. The free tier is a monthly included credit, not an expiring trial. It covers a subset of models with lower per-model rate limits. Purchasing AI Gateway Credits moves your team to the paid tier, and the monthly free credit no longer applies. See Free and paid tiers.

Yes. Purchased credits expire one year after purchase, as stated in the purchase flow.

No. Vercel does not train on your prompts, and AI Gateway itself does not retain prompt or response content: it is deleted once the request completes.

What the upstream provider does with your traffic is a separate question, and AI Gateway gives you governance controls for the provider path, including per-request Zero Data Retention routing, prompt-training disallowance, provider and model allowlists, regional inference, and safety identifiers that isolate end users for provider-side abuse action. Vercel's own retention policy does not constrain providers, so enable these controls when you need a guarantee that covers the whole path. See Security and Compliance for availability and plan requirements per control.

AI Gateway does not log the prompts or responses in your requests. That content is not retained after the request completes. What AI Gateway records is each request's metadata, including its status, model, provider, token usage, cost, duration, authentication method, and every routing attempt. Inspect this metadata in Logs, where routing attempt details are kept for 30 days, or export it to your own observability tool with Trace Drains.

AI Gateway retries a failed request across the model's other eligible providers, and you can configure model fallbacks to reach a different model entirely. Every routing attempt, including recovered failures, appears in Logs, and Uptime and Status explains how provider health is tracked.

Uptime is measured from live traffic and published per model and provider, not promised as a single fleet-wide number. Each model page has Uptime and Status sections, also reachable at vercel.com/ai-gateway/models/<model-id>#uptime, and the endpoints API exposes rolling uptime_last_15m, uptime_last_1h, and uptime_last_1d windows per provider.

AI Gateway reports two different numbers. Provider uptime is the success rate of each provider's attempts. AI Gateway uptime counts only the final attempt of each request, so a request recovered by fallback counts as successful, and AI Gateway uptime can be higher than any single provider's. BYOK requests and 4xx responses are excluded from both.

AI Gateway measures and publishes per-provider latency, throughput, and uptime for every model endpoint. The GET /v1/models/{creator}/{model}/endpoints response includes latency_last_1h, throughput_last_1h, and uptime figures measured through the gateway, so the numbers you see include any routing overhead. For latency-sensitive workloads, fast mode requests the fastest serving path for supported models. See Metrics for how these measurements are made.

The free tier applies lower per-model limits. The paid tier removes them, and a 429 may come from the upstream provider instead. See Rate Limits for the response shape, retry guidance, and the difference between rate limits and budgets.

Open your dashboard's Support entry for technical and billing questions. For custom rate limits, volume discounts, or invoice payment, contact sales.