Jev with Vercel: AI SDK provider and AI Gateway setup

Vercel offers two separate ways to call Jev, TypeSafe’s decision model: an official AI SDK provider package that plugs into the Vercel AI SDK’s experimental_evaluate API, and a listing on Vercel’s AI Gateway that lets you route to Jev by model id without installing anything Jev-specific.

What the integration does

Two different Jev integrations sit under the Vercel name. The first is @ai-sdk/typesafe-ai, a provider package for the Vercel AI SDK. It plugs into the SDK’s experimental_evaluate API, built for typed decisions rather than free text generation. The second is a Jev listing on Vercel’s AI Gateway, at vercel.com/ai-gateway/models/jev, reached by model id typesafe-ai/jev like any other model the Gateway carries.

Both call the same Jev model. The AI SDK provider passes experimental_evaluate a state and a set of Choice, Score or Noul questions, and the SDK builds the request. The AI Gateway uses the same experimental_evaluate function, addressing the model as a plain string instead of a provider object. For teams already routing through the Gateway, that string is the whole integration.

Official support

Both pieces are shipped by Vercel itself, not by TypeSafe. The @ai-sdk/typesafe-ai package is listed as an AI SDK provider on ai-sdk.dev, and the Jev entry on the AI Gateway is a Vercel-hosted listing. No unofficial adapter is needed to reach Jev from a Vercel app.

Setup

1. Install the AI SDK provider with pnpm add @ai-sdk/typesafe-ai, whose npm description reads “AI SDK integration for Typesafe AI.” A different name, @typesafe-ai/ai-sdk, sometimes circulates instead: that path 404s on npm, so avoid it.
2. Import typeSafeAi from the package and experimental_evaluate from ai, then call the model with typeSafeAi.evaluationModel('jev-latest'):

import { typeSafeAi } from '@ai-sdk/typesafe-ai';
import { experimental_evaluate } from 'ai';

const result = await experimental_evaluate({
  model: typeSafeAi.evaluationModel('jev-latest'),
  state: { message: 'Your question here' },
  questions: {
    myChoice: { type: 'choice', criteria: { a: null, b: null } },
    myScore: { type: 'score', criteria: ['Low', 'Medium', 'High'] },
    myBool: { type: 'boolean' }
  }
});

3. Choose question types: Choice (1 to 255 options), Score (2 to 10 ordered levels), or Noul, a boolean question. Every question in a call runs against the same state in one request, and probabilities come back rounded to two decimals.
4. To use the AI Gateway instead, skip the provider package and pass the model id as a plain string:

import { experimental_evaluate as evaluate } from 'ai';

const result = await evaluate({
  model: 'typesafe-ai/jev',
  state: 'The support agent issued a full refund to the customer.',
  questions: {
    refunded: { type: 'boolean', instructions: 'Was a refund issued?' },
  },
})

5. Check pricing before committing to either path. The Gateway page lists $0.042 per million input tokens for typesafe-ai/jev, with a 32,000 token context length. See the speed and pricing guide for what that buys elsewhere, and how to get Jev access for what either path needs first.

What people built with Jev and Vercel

The example projects linked below show the AI SDK provider and the AI Gateway listing in real code, from CLIs to evaluation agents.

Limits and gotchas

The AI Gateway’s own model page says plainly that “detailed capability metadata has not been reported for this model,” Vercel’s own wording. No rate limit, uptime figure, or latency number is published for the Gateway path, and only the input price is confirmed: $0.042 per million tokens for typesafe-ai/jev, with no output price shown, so do not assume it is free.

TypeSafe’s own SDKs, a Python package typesafe-sdk and a JS package @typesafe-ai/sdk, are separate from @ai-sdk/typesafe-ai, the Vercel provider, though all three reach the same Jev model.

Jev is a classification model, an AI model built for decisions rather than conversation, and it only answers the Choice, Score and Noul questions you send with each request. If your app needs open-ended generated text back, neither Vercel path is the right tool. For choosing which model handles a request in the first place, see the guide to intent and model routing.

If you already route traffic through OpenRouter or Cloudflare, a similar model id pattern applies there. See the OpenRouter integration and the Cloudflare integration for those equivalents.

TechCrunch reported on 18 September 2026 that Vercel’s Pranit Sharma said he switched from “OpenAI’s ChatGPT Luna 5.6” to Jev for safety-review classification, citing results 5 to 18 times faster with equal or greater accuracy, his own characterization as reported by a journalist, not an audited benchmark.

FAQ

Does Vercel support Jev officially?

Yes. Both the @ai-sdk/typesafe-ai provider package and the Jev listing on the AI Gateway are shipped by Vercel, not by a third party or by TypeSafe. The provider appears on ai-sdk.dev’s own provider list, and the Gateway entry is hosted directly at vercel.com/ai-gateway/models/jev.

Do I need a TypeSafe API key to use Jev with Vercel?

Not published here. This page only confirms the Vercel-side setup: installing @ai-sdk/typesafe-ai or calling typesafe-ai/jev through the AI Gateway. How authentication works behind either path is not stated in the sources this page draws on, so check the guide to getting Jev access before assuming either route needs no credentials at all.

What does Jev decide inside Vercel’s stack?

Whatever Choice, Score or Noul questions you send it. Jev takes a state, such as a message or a support transcript, and returns typed answers with probabilities, not generated text. Vercel’s Pranit Sharma told TechCrunch he uses it for safety-review classification, switched from an OpenAI model, and reported faster results, though that is his own characterization, not an audited figure.

Is it free to use Jev with Vercel?

No confirmed free tier exists in the sources here. The AI Gateway page shows Jev priced at $0.042 per million input tokens, with no output price visible in that listing. Installing the AI SDK provider itself costs nothing beyond the usual pnpm add, but calling the model through either path bills against that Gateway rate.

Examples for Vercel