Is Jev Free? Every Way to Access Jev, Ranked
Is Jev free? Yes, mostly, and with a deadline. Right now you can run Jev through Vercel's AI Gateway at no charge, but that's promotional pricing that ends on September 25, 2026 — check the model page before you plan around it, because that page is where the date lives. Output tokens are free everywhere, input tokens cost $0.042 per million, and new direct-API accounts reportedly start with $5 in credit. So the honest answer comes in three layers: a time-limited free promo, a small signup credit, and a price structure that's cheap enough that "free vs paid" is almost the wrong question.
What trips people up isn't the price, though. It's the access map. Jev sits in an odd spot — the official API is gated behind a waitlist, while three platforms (Vercel, Cloudflare, OpenRouter) will let you call it today, each with its own quirks and catches. This guide walks all four routes, what each one actually costs, and where "free" quietly stops being free.
If you're new to the model itself — it's TypeSafe's evaluation model that answers typed questions instead of generating text — start with what Jev is first, then come back here.
Is Jev Free? The Three-Layer Answer
Layer one: the promo. Vercel lists Jev at "Free" with the note "Promotional pricing ends on September 25, 2026." That's zero-cost usage through the Gateway — input and output both — for as long as the promo runs. After it ends, pricing reverts to the standard rate.
Layer two: the signup credit. Community reports say new accounts on the direct API get $5 in free credit. I want to be careful here: that figure comes from one source, and the most thorough access guide I found doesn't mention it at all. Treat it as unconfirmed, and let the console be the referee — whatever credit balance you see after signing up is the real number.
Layer three: the structure. Jev charges $0.042 per million input tokens and nothing for output, because output is tiny structured answers — choices, scores, yes/no probabilities — that cost the provider almost nothing to produce. At that rate, $1 buys you roughly 24 million input tokens. Free output doesn't mean a free app, though. Your input adds up, and the code around Jev (preparing state, handling edge cases, generating text when you need prose) carries its own costs.
The four access routes at a glance
| Route | Model ID | Waitlist | Price | Catch |
|---|---|---|---|---|
| Vercel AI Gateway | typesafe-ai/jev | None | Free (promo ends 2026-09-25) | AI SDK only, no usage console |
| Cloudflare Workers AI | typesafe/jev | None | See dashboard | Context capped at 32K |
| OpenRouter | typesafe/jev-1.13 | None | $0.042/M in, $0 out | Missing from the public models API |
| Direct API | jev-latest | Yes | $0.042/M in, $0 out | No published wait time |
My take on choosing: if you write TypeScript, start at Vercel and ship something this week. If you want usage dashboards or you code in something other than TypeScript, queue for the direct API and use OpenRouter in the meantime. Cloudflare makes sense mainly if you already live in its ecosystem — pricing lives in your dashboard, so you'll know within a minute of looking.
Route 1: Vercel AI Gateway — The Fastest Door
A gateway, in plain terms, is a proxy that fronts many AI models behind one API key, so you swap models without rewriting auth. Vercel added Jev on September 16, a day after launch, and by Vercel's own accounting it became the fastest-adopted model in the Gateway's history — within 24 hours it reached nearly 13% of paid teams, about twice the reach of the GPT-5.6 family.
Getting in is refreshingly boring. Create a Gateway key, and call Jev through the AI SDK:
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?',
},
},
});
// result.answers.refunded -> { type: 'boolean', probability: 0.99 }Now the catches, and there are real ones. Jev through the Gateway only works with the AI SDK — the OpenAI- and Anthropic-compatible endpoints that the Gateway also offers don't serve it. If your codebase talks to an OpenAI-style base URL, this door stays shut. And observability is thin: there's no TypeSafe console, no per-key usage page, no rate-limit view. Your Gateway key is the only credential and your own notes are the only ledger. Context through this route is 32K.
Route 2: Cloudflare and OpenRouter — No Waitlist, Fewer Questions
Two more doors skip the waitlist entirely, and both are worth knowing even if you never walk through them.
Cloudflare Workers AI serves Jev as typesafe/jev, no queue, with the same 32K context cap. The model page doesn't publish a price — it points you to the dashboard instead. I like that arrangement less for research and more for honesty: you get the number that applies to your account, not a marketing figure.
OpenRouter lists Jev as typesafe/jev-1.13 (plus a dated build and a typesafe/jev-latest alias), at the same $0.042/M input, $0 output as the direct API. Two things to know. First, Jev doesn't appear in OpenRouter's public models endpoint — that API describes chat-completions models, and Jev's evaluation shape doesn't fit, so it was left out. If your tool builds a model picker by enumerating that endpoint, Jev won't show up; you add it by hand. Second, context here is also 32K, which leads to the strangest discrepancy in this whole access story — more on that in a moment.
Route 3: The Official Waitlist and the Direct API
The direct API is the route TypeSafe itself describes in its launch post: early access, letting developers off the waitlist "as quickly as we can." How long that takes, nobody outside TypeSafe knows — there's no published queue length, no average wait, no end date for early access. What I can tell you is that the waitlist demonstrably moves: independent experimenters were posting results from real early-access keys within days of launch.
Once you're in: sign up at typesafe.ai, and you get console access with keys, usage, and rate limits — the visibility the other routes lack. The endpoint is POST https://api.typesafe.ai/v1/systemone with a Bearer key, and unlike Vercel's route, any language works, not just TypeScript. Model IDs are jev-latest (currently resolving to jev-1.13.0) and jev-preview for the newest build. Rate limits reported in the wild: 250,000 tokens per second and 1,200 requests per minute, though these shift.
Here's that discrepancy. The direct API documents a 64K context window per request — 32K of it reserved for your state plus the longest question. Yet every platform page, all three, lists 32K. The likely reading is that channels expose half the window, but nobody has confirmed that. If your workload carries a large state block, test before you commit to a channel.
While you wait, two things worth installing. TypeSafe ships an agent skill (claude plugin marketplace add typesafe-ai/skills) that teaches coding agents the correct request shape — as one API-tooling blog put it, the skill matters more than the SDK, because a request without a key just 401s anyway. And the official docs are the best single reference for how evaluation actually works.
What Jev Actually Costs Once the Free Runs Out
The sticker: $42 per billion input tokens, which is $0.042 per million, with output free — TypeSafe's line is that output is "too cheap to meter." Numbers from people who've run real things on it:
| Workload | Reported cost |
|---|---|
| Minecraft bot, ~2 minutes, ~150K tokens | ~$0.01 |
| Drone navigation demo | ~$0.10 |
| Browser task: 17 Jev calls, 7 seconds | $0.0039 |
| 500 evaluations through LangChain | $0.34 (vs $28.17 on Claude) |
Treat the first three as directional — they're community-reported, single-source figures. The LangChain comparison is the one I'd anchor on, and its pattern held across runs: not just cheaper, but more stable between repeats.
One economic quirk worth knowing: because Jev prices input rather than output, long documents get relatively cheaper. A traditional LLM reranker's cost grows with how much it writes about each chunk; Jev's cost grows only with what it reads. The longer your chunks, the wider its advantage.
The Boundaries of "Free": What to Check Before You Build
Free access gets you experiments, not guarantees. A few honest boundaries.
That $5 credit is one source's claim, not a promise — verify in the console. Confidence isn't accuracy: a 0.9 probability doesn't mean Jev is right 90% of the time on your data, so calibrate against your own ground truth. The choice list is your ceiling: if the right answer isn't among the options you offered, Jev can't produce it — and when none of your options fit, it still picks one at low confidence rather than refusing. And the dramatic speed and cost multiples on TypeSafe's homepage are vendor-tested, with reference answers generated by other frontier models rather than an independent ground truth. TypeSafe itself flags this.
The sharpest field data point I've seen: a developer ran Jev's checkers against SQuAD 2.0 to test hallucination-killing, and the checkers removed zero — on all eight errors, Jev agreed with the wrong answer at 0.82 to 1.0 confidence. What actually held the line was forcing verbatim quotes plus a plain string check. No model needed. Jev's re-ranking helped; its checking didn't.
None of this makes Jev a bad bet. It makes "free" the wrong thing to evaluate. The question worth asking is whether your workflow fits the evaluation shape — if it does, the pricing almost takes care of itself. And while you wait for direct access, the backstory of TypeSafe, the lab behind Jev, is worth ten minutes.
FAQ
Is Jev still free right now?
Through Vercel's AI Gateway, yes, until September 25, 2026 — that's the promotional end date on the model page. OpenRouter and Cloudflare charge standard rates (input-only, at $0.042/M), and the direct API reportedly opens with $5 of credit. Check the pages before relying on any of this; early access terms move fast.
How much does Jev cost after the promo?
$0.042 per million input tokens, $0 for output, everywhere. Output is free by design — short structured answers are nearly costless to produce — so your bill scales with what you feed in, not what you get back.
How long is the waitlist for direct access?
TypeSafe hasn't said. No queue stats, no averages, no end date — just "as quickly as we can." Anecdotally it moves within days, but the only reliable strategy is to sign up early and use Vercel or OpenRouter in the meantime.
Can I use Jev from Go, Rust, or Python?
On the direct API, yes — it's a plain HTTPS endpoint with a Bearer key, so any language works. Through Vercel you're limited to the AI SDK, which in practice means TypeScript. The OpenRouter listing is the language-agnostic middle path.
What's Jev's context window?
It depends on where you ask. The direct API documents 64K per request; Vercel, Cloudflare, and OpenRouter all list 32K. Unconfirmed, but the channels may expose half the window. Test with your real state size before committing.
Is it worth waiting for direct access?
If you're evaluating Jev seriously, queue now — usage visibility and the higher context window only exist on that route. If you just want to test the waters this week, the Gateway promo is the fastest free access you'll get, deadline included.