Service tiers now available on AI Gateway

**Service tiers** are now available on Vercel’s **AI Gateway**. The feature lets teams choose how each request should be optimized across latency, throughput, and cost, instead of treating every call the same. At launch, tiering is available for **OpenAI** and **Gemini** models.

Mechanically, a tier is a per-request tradeoff setting on the gateway path. A faster tier is aimed at interactive workloads: less queueing and higher token throughput when a user is waiting on the response. A lower-cost tier is aimed at background jobs that can accept more latency in exchange for cheaper execution. The same gateway surface stays in place; the tier selection is what changes how the request is scheduled and priced relative to those goals.

For engineers and builders, that matters because production AI traffic is rarely uniform. Chat UIs, autocomplete, and agent steps that block a UI need tight latency and stable token throughput. Batch summarization, offline evaluation, enrichment jobs, and retry loops usually do not. Service tiers let you encode that split at the request boundary instead of running everything on one default path and paying either too much latency or too much cost.

In market terms, this is Vercel tightening AI Gateway around operational control rather than only model access. OpenAI and Gemini coverage at launch means the first useful split is across two of the most common model families in app stacks. Competing gateway and proxy layers increasingly compete on routing policy, reliability, and cost knobs; per-request tiering is one of those knobs, scoped here to latency, throughput, and cost per request.

Practical takeaway: map traffic by tolerance first. Put user-facing, synchronous calls on the faster tier. Put async and batch work on the lower-cost tier. Watch two things next: whether tier support expands beyond OpenAI and Gemini, and whether your own metrics show that interactive paths actually get lower queueing and higher token throughput while background spend drops without breaking SLAs.