Jev Is Not an LLM, and Vercel Says Nearly 13% of Paid Teams Tried It in a Day. The 444x Is TypeSafe's Own Math.

Every model we have covered this year, from GPT-5.6 to Claude Fable 5, is a text generator we bully into a one-word answer. Route this ticket. Approve this tool call. Is this spam. We write "answer only with JSON" and pay for a paragraph of reasoning we throw away.

Jev, from a San Francisco startup called TypeSafe AI, is the company's first public model, and it is built for that job and nothing else. It launched September 15, 2026. It takes a blob of text (the "state") and a list of typed questions, and returns a choice, a score, or a yes/no for each, with a probability attached. It does not write a word. Vercel put it on AI Gateway on September 16 and, by its own count, nearly 13% of its paid teams had tried it a day later.

The pitch is 193.6x faster and 444.6x cheaper than LLMs. Both numbers are TypeSafe's, against TypeSafe's workflows. Here is what survived the evals page, the failure-modes page, the arXiv paper, and four developer write-ups.

What a Decision Model Actually Does

TypeSafe calls the category "System One Models," after the fast, intuitive System 1 in Daniel Kahneman's *Thinking, Fast and Slow*. The docs: these models "do not write replies, produce code, or generate explanations of their reasoning." You send state (a string, a JSON object, or an array of text) plus questions of three types:

• Choice: one option from a list of up to 255, with a probability per option.
• Score: a position on an ordered rubric, returned as an interpolated value (2.97, say) plus a probability per rung.
• Boolean: yes or no with a probability. Cloudflare's model page calls the type "Noul."

The types mix in one call and are evaluated in parallel against the same state; "adding questions barely changes the response time." Input is text only; images, audio, and video are "not supported (yet)."

CEO Diogo Almeida is the fourth listed author on OpenAI's InstructGPT paper (arXiv 2203.02155, March 2022). TypeSafe's team page says he "co-invented RLHF and InstructGPT"; RLHF predates that paper, so we will say co-author. Co-founders are Erik Gafni (CTO) and Sasha Sheng (COO, ex-Meta FAIR). DCVC led a $40 million seed; Forbes reported a $200 million post-money valuation, relayed here via SiliconANGLE because Forbes blocks fetches.

Almeida, from the press release: "I spent years working on models designed to make AI better at interacting with people. But if AI is going to fundamentally change how work gets done, people can't be the only consumers of intelligence." DCVC's James Hardiman: "TypeSafe is approaching one of the biggest remaining challenges in AI: turning increasingly capable models into technology that developers can reliably build into products at scale."

The sharper line sits on the typesafe.ai homepage, not in the manifesto as some coverage has it: RLHF "creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop." From one of the people who shipped RLHF at scale, that is pointed.

The Seven Jobs Vercel Says It Is For

Vercel's use-case page lists seven: route incoming forms to the right team; prioritize support tickets by reported impact; review agent tool calls before execution; select a response model for each request; categorize documents for search; flag content for moderation review; evaluate generated responses against requirements.

Tool-call approval is the one to watch: LangChain's AutoModeMiddleware "asks for the probability that a tool call is risky or insufficiently authorized" before letting it run, and Langfuse shipped "Jev as a judge" on September 22 behind a feature flag. Vercel itself swapped GPT-5.6 Luna for Jev in its internal safety classification and told TechCrunch it runs 5 to 18 times faster with better accuracy. An anecdote, not a benchmark, but a real production swap.

The AI SDK call, from Vercel's evaluation docs (AI SDK 7.0.105 or later; the API is marked experimental):

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?' },
  },
  providerOptions: { gateway: { zeroDataRetention: true } },
});

console.log(result.answers.refunded);
// { type: 'boolean', probability: 0.99 }

The number you act on is the probability, rounded to two decimals. Armin Ronacher, to TechCrunch: "The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it." Thresholds are your job.

The Multipliers Are TypeSafe's Numbers Against TypeSafe's Workflows

House rules: 193.6x, 444.6x, 238x, and Forbes' "100x" are all TypeSafe figures. To the company's credit, the methodology is public at evals.typesafe.ai and the launch post is candid: "This is where the claims of 193.6x faster, 444.6x cheaper on our home page comes from, and we expect that these are on the higher end of real world gains."

The setup: four workflows (Security Incidents, Agent Trace Observability, Invoice Processing, Customer Service). Reference labels "were generated via an average of the responses of GPT-6 Astra and Claude Fable 5.1, both at high thinking," so the eval measures agreement with two frontier LLMs, not ground truth. Comparison models ran at default reasoning.

TypeSafe's aggregate table, mean of the four workflows:

Read honestly, that is a different headline from the homepage. On TypeSafe's own eval, Jev ties Claude Sonnet 5, beats GPT-5.6 Luna by one point, and trails GPT-5.6 Sol, Claude Opus 5, and GPT-5.6 Terra. Speed and cost are where it wins, by a lot.

Three observations the homepage does not lead with:

• The two headline multipliers come from two different models. From the rounded tables, 78.1 s / 0.4 s is about 195x, the Sonnet 5 time ratio; $0.1761 / $0.0004 is about 440x, the Opus 5 cost ratio. A dev.to analysis reached the same reconstruction and, averaged across all eight comparison setups, gets 97.8x faster and 149.2x cheaper. Still large. Not the banner.
• Against the cheap model, the gap is ordinary. GPT-5.6 Luna costs 8x per case and takes 32x as long, at one point less accuracy. If your triage already runs on Luna, Jev is an upgrade, not a category change.
• TypeSafe flags its own bias. The workflows "were made by individuals on our model capabilities team, so some bias could exist," and the LLM numbers came from OpenRouter, where "there almost certainly is bias here: more complex queries might be routed to better models."

The "238x lower input price than Claude Fable 5.1" line is sound arithmetic ($10 / $0.042 = 238.1 on Anthropic's pricing page), input only, since Jev has no output to compare.

What the Independent Testers Found

Artificial Analysis does not list Jev, so there is no index score. What exists is a week of developer testing, and it is consistent.

Calibration is real but not perfect. A GitHub study (scienthoon/jev-ood-calibration, run September 19 via Vercel Gateway) scored Jev 94.2% on OpenBookQA, 88.1% on CommonsenseQA, and 86.1% on HellaSwag, for about $0.06 total. On 900 synthetic support tickets its expected calibration error was 0.107 against a noise floor of 0.024, and on a question with no knowable answer it was 44.7% accurate at a mean stated probability of 0.74. Calibrated on what it can know; overconfident on what it cannot.

Decomposition beats single questions. Rajesh Beri ran 2,000 phishing emails. One question got Jev 62.6% against Claude Haiku 4.5's 81.3%. Five atomic questions fed into a logistic regression on 1,000 labels flipped it: Jev 95.0%, Haiku 93.2%, at $0.038 per 1,000 emails versus Haiku's $0.462 for one question and $1.02 for five, median latency 239 ms versus 687 ms.

On accuracy, it is a mid-price LLM. An arXiv paper (Ibrahim and Zaki, 2609.24574) tested 7,977 human-labeled items and found Jev 11.6 F1 points behind the best model per task. A 200-item, six-model bench put Jev at 72.5% against Claude Fable 5.1's 84.0%. Speed gains in the dev.to GDE roundup ran 0.5x to 12.1x and cost 0.6x to 478x, depending entirely on the comparator; a single-call triage test found 4 to 7x faster and 31 to 65x cheaper against Claude Sonnet 5, GPT-5.6 Sol, and Gemini 3.8 Flash. The roundup's summary, which we share: level with mid-price LLMs, behind the frontier, with gains far smaller than the banner once the comparator is not a frontier model at 78 seconds a call.

"Zero Hallucinations" Means the Schema, Not the Answer

The homepage says: "Zero Hallucinations - Every Jev decision comes with a confidence estimate, so your software can act when confidence is high and escalate when it is not." True in the narrow sense that Jev cannot return an option that is not on your list. It can return the wrong option with a confident probability. The Register made the point at launch, and akka.io put it plainly: the zero-hallucination figure follows from the output schema and was never measured.

TypeSafe, again to its credit, publishes a "model jaggedness" page for jev-1.13 that opens with "jev-1.13 is fast, calibrated, and good at common-sense judgment but it is not perfect." The failure modes, in its words:

• Literal reading. "Jev answers the question you wrote, not the one you meant. Scoping words, negations, and implied conditions are read at face value."
• Math. "Jev is not a calculator. We strongly recommend implementing any mathematical logic in code."
• Dates. "Jev reads dates as text, not as ordered quantities."
• Large irrelevant state. "Accuracy falls as the state grows with content unrelated to the decision."
• Adversarial content. "Content written to adversarially steer the model, whether that is an injected instruction, a deliberately misleading framing, or text that argues for its own classification, can move the answer."
• Structural invariants, such as complementary probabilities summing to 1, "simply aren't guaranteed by the model."

The adversarial line matters most for tool-call approval, because that is where someone will try. VentureBeat reported an Octomind engineer's test in which one injected instruction dropped Jev's block probability on a risky action from 0.76 to 0.48 and its confidence from 0.64 to 0.22. Pydantic's docs say it straight: "a guard built on Jev belongs alongside deterministic checks, not instead of them."

Pricing and the Fine Print

$0.042 per million input tokens, output free. The same rate is on TypeSafe's docs, the Vercel model page, OpenRouter, and LiteLLM.

Free on Vercel AI Gateway through September 25, tomorrow. The Vercel model page already lists $0.042 per 1M and the changelogs do not state a post-promo rate, so budget on $0.042 and check Friday.

Direct signups are paused. September 15: early access with a waitlist. September 20: "Jev is now available to everyone. No waitlist," with a $5 credit worth roughly 120M tokens. September 22: "We have seen such an immense swell of demand that we have to temporarily pause signups for Jev. We need to ensure quality of service for our existing signups, which will continue to function." The status page logged API latency incidents through this morning. The homepage's "early access" wording is, as of today, accurate again, and the gateways are the no-signup path.

Rate limits move without notice. Direct: 250,000 tokens per second and 1,200 requests per minute, with 429s and a retry-after header, and TypeSafe says limits are "adjusting dynamically" while GPU capacity lands. Vercel applies lower per-model limits on its free tier and warns a 429 may come from upstream. Zero data retention is per-request on Vercel, "Yes" on Cloudflare, and enterprise-only on TypeSafe direct.

Where You Can Run It

The timeline is the story of the week:

Vercel's numbers, from its September 18 post: "Jev reached a tenth of teams within 18 hours. By hour 24, nearly 13% of paid teams were using it. That's 2x the GPT-5.6 family and more than 6x Fable 5.1's share." Two asterisks: this is Vercel's platform data, not an independent count, and it happened during a free promo. Vercel's own closing line: "The next test is whether that early adoption lasts."

Beyond Vercel: Cloudflare Workers AI lists typesafe/jev (no per-token price on the page and no changelog entry, so we cannot date it). OpenRouter has typesafe/jev-1.13; LiteLLM, typesafe/jev-1.13.0; Pydantic AI, typesafe:jev-latest. Testers recommend pinning jev-1.13.0; Vercel's Gateway does not expose the version at all.

For our automation readers: n8n has no official Jev node, only community nodes. The practical path is an HTTP Request node posting state and questions to Vercel's https://ai-gateway.vercel.sh/v1/evaluate, exactly the shape a routing step in a self-hosted n8n wants. Parse answers, branch on the probability, done.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Jev if you have a classification, routing, or scoring step that currently burns an LLM call, and you can write the question as a literal, self-contained decision over a small, relevant state. Ticket triage, form routing, model selection, moderation flags, and LLM-as-judge grading at volume are the sweet spot. Decompose into atomic questions and threshold on the probability; the phishing test shows that is where it goes from losing to Haiku 4.5 to beating it.

Skip it as your only guardrail on agent tool calls, because injected text moves the answer and TypeSafe says so. Skip it for arithmetic, date ordering, or double negatives, and if you need to know why it decided, because it will never tell you. Pin jev-1.13.0; the model is nine days old and the rate limits change without notice.

The frontier is elsewhere: TypeSafe's own table has GPT-5.6 Sol at 74.1% and Claude Opus 5 at 73.1% against Jev's 67.8%, and on the current Artificial Analysis v4.3.2 index the board is led by Claude Opus 5.5 at 58, with GPT-6 Astra and Claude Fable 5.1 at 53. Jev is not on that index at all, and that is the point: it is a different kind of tool. Our model comparison has the current standings side by side.

Sources

• TypeSafe: Introducing System One Models and Jev
• TypeSafe docs: models, pricing, and limits
• TypeSafe docs: model jaggedness, jev-1.13
• TypeSafe evals: methodology and tables
• TypeSafe press release (via Yahoo Finance)
• Vercel: Jev is the fastest-adopted model on AI Gateway
• scienthoon: Jev out-of-distribution calibration study
• dev.to: Jev after eight days of independent tests
• VentureBeat: prompt injection can influence the verdict

*Not sure whether a decision model belongs in your stack or whether you just need a cheaper LLM? Take the free 2-minute quiz and get matched. Then read the frontier half of this week's news: GPT-6 Sol and Luna vs Claude Opus 5.5.*