Tools · · 2 min read

Jev hit 35% of paid teams on Vercel AI Gateway in two weeks. If you call it, `experimental_evaluate` is now `experimental_decide`

Vercel's October AI Gateway Production Index says TypeSafe AI's Jev decision model reached 35.3% of paid teams in its first 14 days but only 0.27% of spend. Here is what Jev returns, what it costs, which model ID to pin, and the AI SDK rename that the older guides still miss.

Vercel’s AI Gateway Production Index for October 2026, published October 15 and covering traffic through September, calls Jev from TypeSafe AI “the fastest-adopted model in AI Gateway history.” It reached 35.3% of paid teams in its first 14 days on the gateway (launch: September 17). Claude Sonnet 5.5 was second at 18.1%, and GPT-6 Luna third at 14.9%. TechCrunch picked up the data on October 16.

Short version: Jev is not a chat model. It returns typed decisions with probabilities, it is very cheap, and a lot of teams tried it. If you are adding it now, use the current AI SDK name, experimental_decide, and pin a version ID if you have tuned thresholds.

Wide reach, tiny bill

The same report shows how small Jev’s share of usage is in money terms:

  • 0.27% of gateway spend during its first two weeks
  • About 1% of September’s tokens and 0.15% of September’s spend
  • The most requests of any model in September, even though it launched halfway through the month
  • 55.6% of week-one teams came back in week two, the fifth-highest return rate among September launches

Vercel’s explanation: a Jev request is “faster, cheaper and uses a fraction of the tokens” of a typical language-model call. The numbers cover only traffic routed through AI Gateway, so they say nothing about direct API usage. TechCrunch also reports that Vercel itself replaced GPT-6 Luna with Jev for classifying harmful commands.

What Jev actually returns

According to Vercel’s Jev and AI SDK guide, you send a state (a string, JSON object or array) plus named questions, and Jev answers each one independently:

  • choice: picks one of up to 255 named options and returns a probability for each
  • score: grades against an ordered rubric of 2 to 10 levels
  • boolean: returns the probability that a statement is true

The guide lists model ID typesafe-ai/jev, $0.042 per 1M input tokens, no output-token charge, and a limit of 64,000 tokens per request (32,000 for state). Zero Data Retention can be set per request with providerOptions.gateway.zeroDataRetention. One gotcha: AI Gateway’s OpenAI-compatible endpoints do not support Jev, so you need the AI SDK or the gateway’s native HTTP API.

The rename the guides miss

Vercel’s guide still says to install AI SDK 7.0.105 or later and import experimental_evaluate. But the ai package changelog for 7.0.128 lists “rename evaluate to decide”. In the current release, 7.0.143, experimental_evaluate survives only as a deprecated alias of experimental_decide. In @ai-sdk/gateway 4.0.116, gateway.evaluationModel() is likewise deprecated in favor of gateway.decisionModel(). Old code still compiles, but new code should use the new names, as the Jev 1.14 changelog examples already do.

Which Jev to pin

Vercel added Jev 1.14 on October 15 as typesafe-ai/jev-1.14.0. Its changelog says the model accepts images as a file part in state and has an experimental long-context mode, providerOptions.typesafe.longContext: true. The raw route is POST https://ai-gateway.vercel.sh/v1/decisions, which uses a predicate question type for yes/no questions.

TypeSafe’s own models page is more conservative. It says Jev 1.14 has been available since October 14, but its alias jev-latest still points to 1.13.0, while jev-preview points to 1.14.0. It lists text-only input for both versions. It also warns that long context, which raises both limits to 262k tokens, is “slower and less accurate.” If you rely on image input or long context, test it on your own data first.

TypeSafe’s advice on aliases is worth following: an alias can move to a new version with no change to your code. If your routing thresholds were tuned on one version, send the versioned ID and upgrade on purpose.

Source: Vercel AI Gateway Production Index, October 2026