Jev free in Vercel AI Gateway until 25 September

What Vercel has made available

The key point is simple: Jev is currently free through Vercel AI Gateway, with the promotional window announced through 25 September 2026. As of 21 September, this is not a new permanent price but a short opportunity to test the model without token charges.

The primary reference is the “Jev API, Pricing & Playground” page in the Vercel AI Gateway catalog. Vercel’s documentation also notes that available models can be discovered through the models page or the models JSON endpoint. In other words, Jev is part of the regular Gateway workflow rather than hidden in a separate demo.

Jev is interesting not as another general-purpose LLM, but as a fast evaluation and routing layer between models. Integration materials list the identifier typesafe-ai/jev; separate guides connect it to the experimental_evaluate flow in AI SDK 7 or newer. That is a concrete use case: quickly deciding which next step or model a workflow needs.

The economics change after the promotion. At the time of the announcement, TypeSafe listed a base price of $0.042 per million input tokens, while output tokens were not billed. Those figures should be kept separate in any estimate: Vercel’s zero price is temporary, while the TypeSafe rate is more useful for planning a long-lived system.

Why this is more than a free toy

The main change is that a routing experiment can be separated from the router’s own budget. A team can test whether a dedicated evaluation layer reduces latency and selects branches more predictably, without blending the result with Jev costs during the promotion.

Free access, however, proves nothing about decision quality. I would first examine routing errors, resilience to ambiguous prompts, and how many expensive calls can actually be avoided. Launch coverage cites latency of roughly 70–500 ms, but that is a stated range rather than a guarantee for a specific workflow.

Systems with frequent short evaluations benefit most: agents, branching pipelines, and multi-model orchestration. The idea that “the router will optimize everything on its own” does not hold up: without a custom trace set and metrics, it merely adds another decision point. After 25 September, the most important question is not the price, but whether the gain persists across the real request distribution.

We previously covered how free access to Pony Alpha on OpenRouter can support low-risk model testing and architecture evaluation. The same approach applies when assessing JEV through Vercel AI Gateway before committing to paid usage.