You read about Jev, went to sign up, and hit a waitlist form. That is where most people stop.
The waitlist is real, but it is only one of the doors, and it is not the one that opens fastest. TypeSafe’s launch post says Jev is “available today in early access” and that the team is “bringing developers off the waitlist as quickly as we can.” Meanwhile a second route has been live since mid-September and asks for no invite at all.
This post ranks the five routes by one question: can you send a real request in the next ten minutes? If you want the model itself explained first, start with what Jev is and come back.
What you are actually getting
Jev is a System One model: you hand it a block of state and a set of closed questions, and it returns calibrated probabilities instead of prose. Three question types cover almost everything you would ask it: noul for yes or no, choice for up to 255 options, and score for 2 to 10 ordered levels.
1. Vercel AI Gateway
This is the fastest door for most readers, and the only one with no waitlist listed. Vercel added Jev to the AI Gateway on September 16, 2026 (changelog). If you already have a Vercel account with Gateway access, you have Jev access. The model id is typesafe-ai/jev.
You call it with experimental_evaluate from the AI SDK, version 7 or later:
import { experimental_evaluate as evaluate } from 'ai';
const result = await evaluate({
model: 'typesafe-ai/jev',
state: 'The support agent issued a full refund of $42.00 and closed the ticket.',
questions: {
refunded: { type: 'boolean', instructions: 'Was a refund issued?' },
},
});
// result.answers.refunded -> { type: 'boolean', probability: 0.99 }
One caveat matters enough to plan around. Vercel’s evaluation docs state that evaluation is available through the AI SDK only, not through the Gateway’s OpenAI-compatible, Anthropic-compatible or Cohere-compatible endpoints. So if your service currently talks to the Gateway by pointing an OpenAI client at a base URL, that client cannot reach Jev. You need the AI SDK in the path.
That rules the Gateway out for a Go service, a Rails app, or anything that reaches models through an OpenAI-shaped SDK. For a TypeScript project already on the AI SDK, it is a one-line model change.
One more thing to know before you lean on this route in production: your Gateway key is the only credential in play, so you get no TypeSafe console, no per-key usage page, and no view of where you sit against Jev’s own limits of 250,000 tokens per second and 1,200 requests per minute. Those numbers are dynamic anyway. Log your own request counts if the volume matters.
2. OpenRouter
Jev is live on OpenRouter with no waitlist. The model ids are typesafe/jev-1.13, the dated build typesafe/jev-1.13-20260917, and typesafe/jev-latest if you want to track the newest release. Pricing matches TypeSafe directly at $0.042 per million input tokens, and the listing shows a 32,000 token context rather than the 64k the direct API documents, so check that ceiling before you move a large-state workload over.
This is the second door that opens today, and it suits you if you already route models through OpenRouter and want Jev billed alongside everything else. One quirk worth knowing: Jev does not appear in OpenRouter’s public model list API, so anything that builds a model picker by enumerating that endpoint will not see it.
3. The official waitlist for the direct API
This is the door everyone finds first, and it is still worth joining even if you take route 1 today. Sign up on typesafe.ai and the console lives at console.typesafe.ai.
Once your invite lands, the shape is a single POST. The endpoint is https://api.typesafe.ai/v1/systemone with a Bearer key. Two model ids are live: jev-latest, which currently resolves to jev-1.13.0, and jev-preview, which points at the newest build whether or not it is an official release. Our guide to creating a Jev API key covers key creation, the first request, and the error codes you will meet.
Be clear-eyed about the cost of waiting: TypeSafe has published no wait time, no queue position, and no target date. “As quickly as we can” is the only commitment on record. Join the list, then build against route 1 while you wait.
Why bother at all if the Gateway works? Direct access gives you every client language, not only TypeScript. It removes a hop from your latency budget, which matters when the vendor’s own end-to-end figure is 70ms to 500ms. And it gives you the console, where keys, usage, and limits are visible in one place. If you want the background on who you are signing up with, we covered TypeSafe AI as a company separately.
4. The agent skill for Claude Code and Codex
TypeSafe ships a drop-in agent skill for coding agents. Install it one of two ways:
claude plugin marketplace add typesafe-ai/skills
# or
npx skills add typesafe-ai/skills
This is not a credential, and it will not get you past a waitlist. What it gets you is a coding agent that builds against Jev correctly. Without it, an agent asked to “add Jev” will reach for a chat-completions call, write a prompt, and try to parse a yes or no out of a paragraph. That is the wrong shape for a model that returns probability fields, and you end up debugging code that was never going to work.
With the skill loaded, your agent knows the request shape, the three question types, and what to do with a confidence value. Pair it with route 1 or route 2, because it still needs credentials behind it. Once the wiring is right, how to use Jev covers the harder part: designing questions the model can actually answer.
5. The official SDKs
TypeSafe publishes a Python SDK and a JavaScript SDK:
pip install typesafe-sdk
Here is the honest version, because several roundups pad their list with this one. An SDK is not a separate way in. It is the most comfortable way to use route 1 or route 2 once you already hold credentials. Installing the package on a machine with no key gets you an import statement and a 401.
Use them anyway once you are through. They handle retries, typed responses, and the probability fields, which saves you writing the same parsing twice.
6. Run a Jev-like model locally
If you are blocked entirely and want to prototype the pattern this week, the community has shipped reproductions: OpenJev, mini-jev, jevlike, and an MLX space for Apple Silicon. They let you build the calling code, the question set, and the threshold logic now, then swap the endpoint when real access lands.
None of them is Jev. None reproduces its training, and calibration is exactly the property that is hardest to copy. Treat them as a scaffold, not a substitute. We compared them in detail in OpenJev and the open-source Jev alternatives.
Which route should you pick
- Fastest today: Vercel AI Gateway, if your code is TypeScript and can call the AI SDK directly. Minutes, not days.
- Best long term: the direct API. Any language, one less hop, and the console is where rate limits and usage actually live. Join the waitlist now so the clock starts.
- Best if you are blocked entirely: a local reproduction for the prototype, plus the agent skill so whatever your agent writes has the right shape when the real key arrives.
- Not a route on its own: the SDKs. Install one after you have credentials, not before.
Build the call once in Apidog, point it at either door
Here is the practical reason to set this up in an API client instead of curl. Whichever door you come through, the request body is the same shape: a state string and a questions object. So build the call once in Apidog, store the key as an environment variable, and you can point the same saved request at the TypeSafe endpoint or your Gateway route by switching environments. No second collection, no copy-paste drift between the two.
Then add an assertion on a probability field, something like answers.refunded.probability > 0.9 for a case you know is true. That single assertion tells you immediately whether a new route behaves the same as the old one. When your waitlist invite lands and you flip from the Gateway to the direct API, you run the saved request against the new environment and read one green check instead of eyeballing two JSON blobs. Download Apidog and the setup takes about as long as reading this section.
FAQ
Why doesn’t Jev show up in OpenRouter’s model API?
It is listed on the site but missing from the public /api/v1/models response, which returns 446 models with no Jev entry. That endpoint describes chat-completions models, and Jev is an evaluation model, so it does not fit the shape. If you enumerate models programmatically to build a picker, Jev will not appear and you will have to add it by hand.
How long is the waitlist?
Unpublished. TypeSafe has not shared a queue length, an average wait, or a date when early access ends. The only public statement is that the team is moving developers off the list as quickly as it can. Plan as if the invite might take weeks, and build on the Gateway in the meantime.
Does going through the Gateway cost more?
Jev is priced at $0.042 per million input tokens, and output tokens are not charged, which makes it cheap either way. The Gateway adds its own billing layer on top, so compare your Vercel invoice against the console before you move production traffic. Context is 64k per request, with 32k of that reserved for state plus the longest question, and our guide to using Jev covers how to stay inside that budget.
Can I use Jev from Go, Ruby, or Rust?
Only through the direct API, which means route 2. The Gateway path needs the AI SDK, and TypeSafe publishes SDKs for Python and JavaScript only. From any other language you send a plain POST to https://api.typesafe.ai/v1/systemone with your own HTTP client once your key is live.
Is the vendor’s speed claim reliable?
Treat it as a vendor claim. TypeSafe reports 70ms to 500ms end to end, and 193.6x faster and 444.6x cheaper on its own tested workflows. Those are the company’s numbers on the company’s workloads, not independent benchmarks. Measure your own latency on your own state sizes before you promise anything to a product team.
Start with the door that is already open
Two of these five routes can be running today: the Gateway if you write TypeScript, a local reproduction if you do not. The waitlist is worth joining on the same afternoon, because it is the only route that is language-agnostic and the only one where you control the key.
Whichever way you get in, build the request once in Apidog, assert on a probability field, and keep both environments side by side. Then the day your invite lands is a config change, not a rewrite.
