Jev

Out of six requests users bring to Jev, it handles only one without caveats: checking the price. You cannot try it for free, and no key is available today.

Your code already asks an LLM “which of the three departments,” “spam or not,” “how angry is the customer.” Jev handles such questions in a third of a second for fractions of a cent; see the Norwegian documents example in chapter 2.

You need hundreds of thousands of small decisions an hour: moderation, sorting, scoring. Checking nine thousand product pairs cost $0.32 (paddo.dev, Sep 19, 2026).

You value the same answer to the same question. On public-procurement data, Jev repeated its decision in 98% of cases, while gpt-oss changed its answer on every sixth item (Habr, Sep 21, 2026).

You need text, code, or an answer based on an image. Jev accepts text only, and it guessed SVG drawings two and a half times less accurately than Claude Sonnet 5 using actual images (Sep 19, 2026).

You want to ask one large question such as “can this code be deployed.” Almeida himself says this will work poorly: the model excels at small questions that combine into more complex tasks (Sep 17, 2026).

You do calculations or compare dates without checking. The vendor says the model does not calculate and reads dates as text; English is its primary language.

How to get a key and call the model, in chapter 3; how it differs from GPT-5 nano and Haiku, in the comparison. If you are just starting with AI, open the vibe coding introduction.

Three question types. Jev understands three question forms. Choice selects one item from a list, Score gives a score on a scale, and Noul answers “yes” or “no” with a number from 0 to 1. You can ask dozens of questions about one text at once; the model processes them in parallel, so response time barely increases.

Racing an LLM. In its launch video, one support ticket with three dozen questions was given to Jev and GPT-5.6 Terra. Jev answered in a tenth of a second; the LLM was still waiting for its first token. The vendor notes that short text favors its model (Sep 15, 2026).

A judge for an agent. LangChain engineers Daniel Shi and Sean Roche ran the same weather-agent answers hundreds of times through four judges and checked them against human labels. Jev matched the human in every case, Claude Sonnet 4.6 was wrong one time in five, and the run cost 80 times less (our news story, Sep 20, 2026).

Norwegian hearings. Emil Lindfors gave Jev two dozen documents on resource rent in salmon farming and asked it to identify each author's position. The model got the position right five times in six and answered per document in a third of a second; DeepSeek without reasoning took almost three. More detailed question wording, to the author's surprise, made it less confident (lindfors.no, Sep 18, 2026).

Product catalog. On blog.r6i.it, product classification through a category tree was moved from an agentic GPT-5.2 loop to Jev. A product moved through the tree seven times faster, with half as many model calls. Paths matched seven times in ten, and a Jev error on the first branch is irreversible: it does not go back up the tree (Sep 21, 2026).

Where it stumbled. Bartosz Mikulski converted four hundred Quick, Draw! sketches into SVG text and asked it to guess the object. Jev guessed one in three and called more than half the sketches airplanes; Claude Sonnet 5, using the actual images, guessed nine out of ten. The model cannot see images at all (Sep 19, 2026).

More. On Hacker News, a user named silbercue guided Jev through websites as a browser agent: the model chose where to click and made no mistakes in six scenarios, while a separate small model typed text into fields.

Our experience. We had Jev re-evaluate two hundred fresh posts from our news feed and compared the results with our large-model judge. They matched in three out of four cases, and the whole run cost three cents. Where Jev is unsure, the judge is usually unsure too; see the breakdown in the article (Sep 21, 2026).

Waitlist. Registration at TypeSafe is closed: on typesafe.ai you submit an application, and a key is issued after an invitation in the console.typesafe.ai console. The documentation lists no free tier or starter credits.

Call. One endpoint, POST https://api.typesafe.ai/v1/systemone, a key header, and a two-field body. state holds the text; questions holds a list of questions with options. SDKs are available for Python (typesafe_sdk) and JavaScript (@typesafe-ai/sdk); the default model is jev-latest, currently jev-1.13.0.

Through gateways. Without a waitlist, the model is available through Vercel AI Gateway as typesafe-ai/jev, and through Cloudflare AI and OpenRouter as typesafe/jev-1.13, at the same price. At Vercel, this is not a regular chat call but a separate evaluate function in AI SDK 7, because Jev has its own request format.

For agents. TypeSafe has released a skill for coding agents. In Claude Code, install it with claude plugin install typesafe@typesafe-ai; it teaches the agent how to write questions for the model.

Limits. 64 thousand tokens per request and 32 thousand for the text plus the longest question, 1,200 requests per minute, and 250 thousand tokens per second. The vendor warns that limits fluctuate and may change without notice while demand exceeds capacity.

Vendor measurement. On its four real-world tasks, TypeSafe measured Jev at 67.8% accuracy for $0,0004 and 0.4 s per case. GPT-5.6 Sol scored 74.1%, Claude Opus 5 73.1%, GPT-5.6 Terra 67.9% at $0.03 and 10 s. The benchmark for correct answers was the averaged responses of GPT-6 Astra and Claude Fable 5.1, rather than human labels (evals.typesafe.ai, Sep 21, 2026).

Where 193.6 and 444.6 come from. The vendor derives these multipliers on its homepage from the same measurement and calls them an upper bound. By our recalculation of its table, 193.6 comes from dividing Claude Sonnet 5's 78.1 s by Jev's 0.4 s, and 444.6 from dividing Opus 5's $0,176 by $0,0004. Speed is compared with the slowest model, cost with the most expensive.

Independent runs. Three measurements from 18–20 September agree: versus low-cost LLMs, the speed gain is two to three times. On Banking77 with 77 classes, Jev scored 83.2%, GPT-5.4 nano 79.3%, GPT-5.6 Terra 87.5%, and a small classifier trained on these data 93.3% (ickma2311, Sep 18, 2026). On the full Banking77 test of 3,080 examples, JevBench got 80.3%, and on a phishing set measured 239 ms latency from France versus 687 ms for Claude Haiku 4.5 (Sep 20, 2026).

It depends on the question. On phishing emails, Jev with one question scored 62.6% versus 81.3% for Haiku 4.5, but after splitting it into five questions with tuned weights, 95.0% (XenoSpectrum, verified Sep 21, 2026). A simple regular expression scores 91.8% on the same set.

Token price. $0,042 per million input tokens; output tokens are free. In its documentation, the vendor counts in billions: $42 per billion. There is no separate pricing page, plans and minimum payment are not stated, and the price appears only in the model description (checked Sep 21, 2026).

Task price. A thousand hearing documents cost $0.22 versus $1.31 for DeepSeek without reasoning (lindfors.no, Sep 18, 2026). Classifying 1,018 articles cost $0.08 (Habr, Sep 19, 2026). A Doom bot made 10 requests per second for $7 an hour; this is the vendor’s only real-world figure.

The price catch. In its blog, TypeSafe writes that it cannot prove the absence of a subsidy and expects a long period to confirm the price is sustainable. Its homepage, meanwhile, says the model is served profitably. Limits change without notice.

Format, not truth. Jev cannot return an option outside your list, so the “zero hallucinations” claim on the homepage is true in that sense. It can choose the wrong option from the list; the vendor says so directly: the model guarantees the response format, not the correctness of every decision.

What confidence means. Calibration is promised across answer groups: of decisions with a 0.8 probability, around 80% should be correct. It guarantees nothing about an individual answer. In ickma2311’s run, Jev’s confidence did not separate its errors better than a standard LLM’s verbal confidence (Sep 18, 2026). The vendor has not published a paper, calibration curve, or calibration-error figures; our analysis of this claim appeared in the feed on 21.09.2026.

Nine weak spots. TypeSafe maintains its own failure page for jev-1.13: literal reading of the question, arithmetic, date comparisons, references across several steps, long text with noise, adversarial insertions, conflicting criteria, rules that must align across answers, and generation. The vendor has one recommendation: calculate and compare in code, leaving judgment to the model (checked Sep 17, 2026).

Data. According to the documentation, the model is not trained on customer requests, and enterprise customers have a no-data-retention mode. Jev cannot be fine-tuned on your data; the same weights serve every account.

Who is behind it. Diogo Almeida is the fourth of twenty authors of InstructGPT, the 2022 paper that led to ChatGPT, and one of around fifty contributors to the ChatGPT launch blog post. The phrase “co-inventor of ChatGPT” from his tweet does not appear in OpenAI documents. The company was founded in San Francisco in 2024 and received $40 million from DCVC on 15.09.2026.

Jev or GPT-5 nano. Nano can write text and costs $0.05 for input and $0.40 for output. On Banking77, Jev was four points more accurate and twice as fast per call, 0.43 s versus 0.9 s (ickma2311, Sep 18, 2026). If you need one decision and no text, Jev is cheaper; if you also need to write a customer reply, nano handles both tasks.

Jev or Claude Haiku 4.5. Haiku understands images, writes text, and costs $1 for input. On phishing with one question, Haiku scored 81.3% versus Jev’s 62.6%; after breaking it down into five questions, Jev reached 95.0% (XenoSpectrum, checked Sep 21, 2026). Jev’s win here comes at the cost of manual work on the questions.

Jev or your own classifier. A small model trained on the same Banking77 data scored 93.3% versus Jev’s 83.2% and responded in 0.01 s (ickma2311, Sep 18, 2026). But such a model must be trained for every task, while Jev takes the task from the question text without training. On Hacker News, it was called more of a classifier than a language model, and the founder replied “very accurately.”

Where each call fits. Jev does not fit the usual chat format, so it cannot be plugged into OpenAI-compatible libraries with a single line. Nano and Haiku work through the familiar OpenAI and Anthropic APIs.

Three days passed from stealth launch to the first independent measurements. Everything below is dated by the authors' publications.