Jev, the decision-making AI model developed by TypeSafe AI, reached 35.3% of paid teams on Vercel AI Gateway within its first 14 days, according to Vercel’s latest usage figures. TechCrunch reported the milestone on Friday, confirming that Jev achieved the highest early team adoption of any model in the gateway’s history.
The headline is striking, but the spending figures tell a more useful story for developers. Jev accounted for just 0.27% of gateway spending during its first two weeks. It is attracting a large number of teams without generating inference costs like general-purpose language models. Here is what Vercel actually measured, what Jev does, and how to test the model in an application.
Jev’s adoption figures: what Vercel actually measured
Vercel published its October AI Gateway Production Index on 15 October. It compared September model launches using three separate measures: how many paid teams tried a model, how much gateway spending it attracted, and how many first-week teams returned in week two.
| Model | Reach in first 14 days | Week-two return rate | Spend share in launch window |
|---|---|---|---|
| TypeSafe AI Jev | 35.3% | 55.6% | 0.27% |
| Claude Sonnet 5.5 | 18.1% | 60.5% | 3.21% |
| GPT-6 Luna | 14.9% | 64.8% | 0.88% |
These figures come from Vercel’s October 2026 AI Gateway Production Index. The comparison uses external paid teams using AI Gateway, not every Vercel customer, every AI developer or users accessing the same models directly through their creators. Vercel defines reach against paid teams active from the beginning of the launch month through its data cut-off. The return rate means a team used the model in both weeks; it does not measure accuracy, customer satisfaction, or long-term retention.
Jev also received more gateway requests than any other model during September, despite arriving midway through the month. Yet it represented only 1% of September’s gateway token volume and 0.15% of that month’s estimated spend. The 0.15% monthly figure and the 0.27% first-two-weeks figure use different time windows and should not be confused.
Another detail is missing from the fastest-adopted headline: Jev ranked sixth on Vercel’s combined September launch scorecard, which also considers spending and return rates. Wide experimentation is clear. Durable leadership across every measure is not.
What is Jev, and who actually makes it?
Jev is built by TypeSafe AI, not Vercel. The company, co-founded by former OpenAI researcher Diogo Almeida, calls it a System One decision model. Vercel distributes access through AI Gateway, alongside models from other providers.
Unlike a conventional large language model, Jev doesn’t draft answers, write code, or generate paragraphs. An application supplies relevant state and predefined questions, then receives typed results. Supported question types include Boolean probability, a choice from named options, and a score on a defined scale. Multiple questions can be evaluated in one request.
Consider an AI support workflow. A generative model may be needed to write a helpful reply, but the preceding steps are narrower: Does the message request a refund? Which queue owns it? Does it require human review? Jev can make those classifications without paying for a generated explanation that the software must subsequently parse.
That doesn’t mean Jev can’t make mistakes. Returning a probability instead of prose removes a text-generation and parsing step, but a confidently wrong classification can still trigger the wrong action. Deterministic business rules, such as eligibility for a refund, should remain in application code.
Jev 1.14 is available on Vercel AI Gateway, with image input
Jev launched in September, so this isn’t a new model release on 16 October. Vercel listed the original model in September and announced Jev 1.14 on 15 October 2026. The newer version adds images as part of the decision state and an experimental long-context option for larger documents, logs and histories. Image input lets the model assess a supplied image; it does not turn Jev into an image generator. Developers can add an image file to the state parts or enable experimental long context with providerOptions.typesafe.longContext: true.
| Developer detail | Current Vercel listing |
|---|---|
| Provider and model | TypeSafe AI, typesafe-ai/jev-1.14.0 |
| Unversioned model alias | typesafe-ai/jev |
| Listed standard context | 32K tokens |
| Listed input price | $0.04 per million input tokens |
| Output-token price | No separate output-token rate shown |
| Access | Vercel AI Gateway via AI SDK, supported HTTP API or TypeSafe client |
At the listed rate, ten million billable input tokens would cost $0.40 before any other applicable service charges. This is a price calculation, not a DIY AI performance or cost test. Check the live model listing and any additional gateway or account charges before committing to a production budget. Experimental long-context behaviour also needs testing rather than assuming the standard 32K limit has disappeared.
For a new TypeScript integration, configure AI Gateway credentials on the server, use AI SDK 7.0.128 or later, and call the current experimental_decide API. Older examples using experimental_evaluate refer to a deprecated name. A minimal decision request looks like this:
import { experimental_decide as decide } from 'ai';
const result = await decide({
model: 'typesafe-ai/jev-1.14.0',
state: 'The customer reports a cracked screen on delivery.',
questions: {
damaged: {
type: 'boolean',
instructions: 'Does the customer report delivery damage?',
},
},
});
console.log(result.answers.damaged.probability);The result is a probability that the stated condition is true, not permission to issue a refund. Developers using a direct HTTP integration can also access Jev through Gateway’s /v1/evaluate endpoint. Gateway requests appear in its usage, budget and reporting tools; teams handling sensitive state should explicitly configure the appropriate no-training or zero-data-retention controls.
The deployment risk is not JSON parsing: it is trusting the wrong decision
Early developers experimenting with Jev for routing and multilingual classification have encountered a recurring practical issue: the model cannot account for business context the application doesn’t supply. A message might look like a fresh support request even when another specialist already owns the case. A confidence score does not restore that missing fact, and occasional high-confidence mistakes still need handling.
Before replacing a working classifier or lightweight language-model call, teams should take five steps:
- Identify the actual decision. Specify the allowed outcomes and the action each one would trigger. Keep open-ended writing with a generative model.
- Build a labelled test set. Include routine requests, ambiguous wording, multilingual inputs and costly false positives, not only clean examples from the documentation.
- Measure the complete workflow. Compare end-to-end latency, accuracy, review workload and cost per accepted decision, including retries and fallback calls.
- Calibrate the hand-off. Route uncertain cases to a stronger model or a person. Do not copy an arbitrary confidence threshold from another application’s tests.
- Pin and monitor the model version. Track false decisions, input mix and calibration over time. Test Jev 1.14 separately from the unversioned alias before changing production traffic.
This approach is especially useful for agent workflows that repeatedly decide which tool to use or whether to proceed with an action. If an agent is connected to external services through MCP, the model’s proposed action and the tool’s execution permissions need separate controls. Our OpenCode MCP setup guide covers that permissions and integration layer.
What Jev’s rapid adoption means for developers
Jev’s immediate appeal is economic and architectural. Instead of using an increasingly capable language model for every small branch in a workflow, developers can reserve language generation for tasks that genuinely require it. High-volume classification, prioritisation and agent routing are credible places to test a specialist decision model.
The adoption chart has limits to the inference we can draw. Vercel AI Gateway is a particular distribution channel, not a census of the AI market. Cheap calls can generate high request counts without consuming much of a team’s budget. Nor do the figures establish that Jev is more accurate than alternatives on a specific dataset. Our AI model cost comparison explains why price per successful task is more informative than an advertised token rate.
For now, Jev’s 35.3% reach is convincing evidence of developer interest in typed decision models. The more meaningful next test is whether those teams keep using it, whether Jev 1.14 improves their real workloads and whether the savings survive the cost of reviewing incorrect decisions. Teams with a measurable routing or classification bottleneck have good reason to trial it; teams seeking a better writing or coding model should look elsewhere.
