Decide who pays for AI Gateway before developers ship plain model strings, because the AI SDK sends those through Vercel by default. Then set budgets at every scope, decide where bring-your-own-key fits, and switch on the data and provider controls your security review requires.
When you need this
- Developers are calling models with plain strings such as openai/gpt-5 and you have not decided who pays.
- Finance wants spend capped per team, project, key, or person.
- Security requires zero data retention or a fixed list of approved providers.
Evaluation criteria
Know what a plain string does
Where does a model id without a provider package go?
Look forAI Gateway is the default global provider. A string model id routes through it with an AI Gateway API key or, on Vercel deployments, an OIDC token. Import provider packages such as @ai-sdk/anthropic with your own keys when calls must go straight to the provider.
Understand the tiers
What does the free tier cover, and what changes when you buy credits?
Look forEvery Vercel team gets US$5 of monthly gateway credit on a subset of models, with lower rate limits. Buying credits ends the monthly free credit and opens the paid tier to every model and bring-your-own-key. Purchased credits expire after one year. Both tiers charge provider list prices without markup.
Set budgets at every scope
Who can spend how much, and what happens at the limit?
Look forSet budgets for teams, project OIDC traffic, API keys, or team members. Choose daily, weekly, or monthly resets, plus defaults for new keys and projects. Each request checks a soft cap at startup, so final spend can exceed it. Enable alerts at 50, 75, and 100%, which default to off.
Decide on bring-your-own-key
Should requests use your provider contracts?
Look forBring-your-own-key adds no fee but needs purchased credits, and its spend counts toward no budget. When a request with your key fails, the gateway retries with Vercel's credentials and charges your credits, so watch that fallback if contracts require your own accounts.
Turn on data controls
Which providers see your prompts, and do they keep them?
Look forZero Data Retention on a single request is free on Pro and Enterprise, and turning it on team-wide costs US$0.10 per 1,000 requests. A team-wide provider allowlist costs the same per 1,000 successful requests. Filtering providers on one request with the only option is free. One developer used it to keep OpenAI model calls off Azure-hosted deployments after errors.
Price reporting and traces
Which add-ons bill outside the token price?
Look forCustom reporting costs US$0.075 per 1,000 tag, user, or quota writes and US$5 per 1,000 queries. Trace drains cost US$0.05 per 1,000 traces and US$0.50 per GB, billed through your Vercel plan's Drains usage with no allowance on Pro.
Reconcile spend
Does tracked spend match what you expect?
Look forLogs show each request's cost. The credits and generation endpoints return the balance and per-request cost, and the AI SDK response carries the gateway cost in provider metadata. One developer's structured workout generations cost US$0.03 to US$0.05 each through the gateway.
Questions to ask vendors
- Which teams send traffic through AI Gateway, and which go direct with our own keys?
- What budgets and alert thresholds apply to each team, project, and key?
- Do our provider contracts require bring-your-own-key, and how do we watch the fallback to Vercel's credentials?
- Do we need Zero Data Retention or a provider allowlist team-wide?
- Which add-ons are switched on, and who reviews their charges?
Red flags
- Plain model strings in production with no decision about the gateway.
- No budgets, or budgets with no alerts.
- Bring-your-own-key spend assumed to be capped by gateway budgets.
- Trace drains at full sampling with no estimate of their cost.
Products in this piece
Sources
- Official documentation
- 6
- Public source
- 1
- AI SDK docs: Choosing a provider (AI Gateway as the default global provider)
- Vercel docs: AI Gateway
- Vercel AI Gateway product page
- Vercel docs: AI Gateway pricing
- Vercel docs: AI Gateway budgets and spend limits
- Vercel docs: AI Gateway FAQ
- Adam Rackis, Master.dev blog: Having fun with Vercel's AI SDK and AI Gateway (Aug 31, 2026)