The fastest-adopted AI model can't write a sentence. Here's what it's for.

On Sep 18, 2026, Vercel published its AI Gateway numbers for a model called Jev, from a company called TypeSafe. Within 24 hours of listing, nearly 13% of the gateway's paid teams had used it. That's twice the first-day share of the GPT-5.6 family and more than six times Claude Fable 5.1's.^[1] Those are two of the most capable general-purpose models there are. The one that beat both of them to a team's first request can't produce a word of prose.

The other big gateway tells the same story from another angle. On OpenRouter, by its fourth full day, Jev was the most-requested of 556 models, with 82 million requests on Sep 21 alone, while sitting only thirteenth by tokens.^[2] Lots of requests with only a few words in each is what decisions look like in traffic. A chat model's traffic runs the other way.

So I put it on the bench: two dozen experiments on public data across four real jobs, plus a small working demo, between Sep 19 and 21, 2026. This page is what it's good for, where it lost, and why the price matters more than the speed.

Here's what you'd likely assume about a cheap model that only answers closed questions, and what the bench said about each.

The answer in one breath

Four real jobs from this site's research instrument, run on public data, and one demo built to show the mechanism. Every scored comparison used a fresh sample of 118 to 154 items that no tuning had touched, graded against blind reference labels.^[5]

You don't chat with it. You hand it a form. You send the thing to look at (a message, a ticket, a record, any text) and questions you wrote yourself, each with its allowed answers. It returns an answer to every question, with the odds attached.^[4]

Because the answer can only be one of your options, it can't invent a category, wander off topic, or pad a reply. What comes back is a number your own code can act on: keep it, drop it, sort by it, or send the item to something smarter. There are three shapes of question, and everything on this page is built from them.

The name is a clue. Jev is short for William Stanley Jevons, the economist who noticed that as steam engines got more efficient, Britain burned more coal, not less. Cheaper work meant more work was worth doing. TypeSafe says that's exactly why it chose the name.^[4] Make a judgment cheap and people start making it everywhere. At $0.042 per million tokens read, with nothing charged for the answer, a paragraph costs a few millionths of a dollar.^[2]

A custom predictive model can answer fuzzy questions, like whether a customer is about to leave or where a ticket should go. But it needs historical data, labeling, someone to train and check it, and upkeep when the world drifts. So a classifier had to be worth enough to pay for its own construction, and most questions never made the cut. A chat model brought the effort down to an afternoon, but for a bounded question it's a lot of machinery: slow, relatively expensive, and producing prose that has to be parsed and checked.

A decision model sits in the gap: messy text in, a bounded answer with odds out, at the speed of ordinary application code and at a price closer to a database lookup than to a model call.

It works alongside code, predictive models and chat models, and connects them. The model turns meaning into numbers, and ordinary software sorts, filters and routes on those numbers. Parts C and D are the two shapes I measured: decision model then plain code, and decision model in front of an expensive model.

The demo is a chat desk for a made-up airline, NorseStar Air, and you play the customer. Behind the chat sit the things the airline could offer (a cheaper fare, a better seat, a miles upgrade, a free checked bag, a nonstop), which combine into 770 buyable packages. A short scoring formula ranks every one of them. It's ordinary code, eligibility rules and a few weights, with no model in it. It knows the customer's status, miles balance and usual cabin. In the first version it couldn't understand a word the customer typed.

One customer, one sentence: "I'm flying for work and don't want to use my miles. I have a large suitcase I need to bring and I need my leg room. What's the best deal I can get for a direct flight?"

Switched offSwitched on

This site publishes rankings of public code repositories, and by policy doesn't name tools built for attacking systems. That's ten thousand repositories, refreshed constantly, with a judgment call on each. Before this, an expensive reader model read every one. A later build asked whether a free model on a Mac mini could take that first look instead.

The demo is public, and so is its code. A bring-your-own-key build runs entirely in your browser: paste one OpenRouter key, and your browser calls the decision model and the chat model itself. Nothing is stored anywhere. Click any recommendation to see every term that built its score. It's at jevdemo.worthington.cloud.^[6]

1. Write the decision down before you pick a model.

   If a careful person holding the written rule could answer in a second, a decision model is a candidate. If they'd have to read, weigh and decide, it isn't.

   From f10

2. Choose the question type before you write the question.

   A yes-or-no, a pick-one and a rating are different tools. Keep a long option list flat.

   From f11, f12

3. Put it at the front of the pipeline, set to be wrong in one direction only.

   Set the cut-off by what you can afford to miss, and let something that can read handle what survives.

   From f6, f7

4. Use it as the second opinion, wired clumsily on purpose.

   It's worth buying because it fails somewhere else.

   From f8

5. Let it turn sentences into numbers, and let readable code decide.

   Show the decision beside the reply, and read the customer again on every turn.

   From f1, f2, f3

6. Write the question you might not need.

   It only hears what you asked about, and a rating's odds work best as thresholds.

   From f4, f5

7. Spend the budget on the written rules and on scoring, not on the model.

   Have two careful readers label 150 of your own items blind first. If they don't agree, no model will.

   From f13

8. Before you believe an accuracy number, ask who wrote the answer key and how each model was called.

   Check that the confidence sorts right from wrong, score once on fresh data, and hold a vendor's 445× to the same test.

   From f15

Under other names the winning shape is everywhere: screening documents against criteria, routing a ticket to a queue, checking that a quoted passage supports the claim citing it, catching a prompt-injection attempt before an expensive model reads it, deciding whether a conversation needs a person. The losing shape looks almost identical from the outside: the same text, the same list of labels, and a rule that takes a paragraph of judgment to apply.

One piece of standing advice backfires if read loosely. "Combine in code" means combining separate factors, not re-applying rules the options already state. Doing both cost four to five points every time.^[5]

• That these scores are accuracy. They're agreement with a reference reader that is itself a model.
• That any of it holds on insurance claims, support tickets or medical records. It's one kind of public text, from one bench, over three days.
• That the economics are measured. The ladder, the falling floor and the Jevons reading are how I read the numbers.
• That fast adoption means lasting use. The adoption figures are one gateway's first day, and a model can be dropped as fast as it's picked up.
• That it's safe on hostile input. Text written to steer a classifier can steer it.
• That it's private. It's an outside service, so anything sent to it leaves your machine.

No guesses were registered for this build. What was written down before any data was drawn was the method: the rule for each job, saved to a file first; reference labels written to files by a reader that saw no other model's answers; and a design frozen and scored once on a sample the tuning never saw.^[5] Here's how each held.

Every idea tried on the bench, and what became of it.

The same mistake happened twice in one evening, in two unrelated experiments. Both times a headline figure was a model measured against something it had produced itself. Both times the number flattered it, and neither was caught by review.

Fifteen points of its twenty-four-point lead were how it was asked, not which model answered. And underneath it, the answer key for that job had been written by the same model working the same careful way, so 96.3% was never scored against an independent standard. It was scored against its own house style. Four hours earlier, on the other job, the answer key had matched the cheap model's own verdict on 99.6% of rows, 10,672 of 10,713.^[5]

It's worth holding the vendor's benchmark to the same two rules. Its reference labels were averaged from two frontier chat models, and its workflows were written by its own team.^{[3, 4]} That doesn't make the numbers wrong. It makes them the vendor's numbers. One news roundup of published field measurements tallied real-world speedups of 1.16 to 6 times.^[7]

• Agreement, not accuracy. Every score is how often the model matched a reference that is itself a model.
• Small samples. 118 to 154 items for every scored comparison, plus two full passes over the 10,714 repositories. A 2 to 3 point gap is noise.
• One corpus, one domain, one bench. Public code repositories and research abstracts, Sep 19 to 21, 2026, one operator.
• The demo is an illustration. Its customers and airline are made up, and its screens are one turn for one customer. Its decision-model cost is an estimate from token counts at list price.
• Speed for a live product was measured once, lightly. The experiments ran in batches. The demo is the only live measurement: a handful of turns from one laptop, landing around 0.6 seconds for twenty-four questions. The earlier thirteen-question build answered in about half that. The questions are cheap, not free. No load, no concurrency, one network.
• Hostile input and privacy weren't tested. Both ruled out some of the most attractive uses before they were tried.

Personal time, personal hardware, public data. Two dozen numbered experiments ran on Sep 19, 2026, each with its sample size, its random seed, its cost and the file holding every individual answer.^[5] The model was typesafe/jev-1.13 on OpenRouter; the dated build that answered was typesafe/jev-1.13-20260917. The demo was built the next day on the same laptop, then rebuilt on Sep 21 to ask 24 questions instead of 13 and rank 770 packages instead of 10 offers. The chat model in the demo is Claude Haiku 4.5.^[6]

First build, switched on

The video at the top was made with the video kit used for every explainer on this site. Claude wrote and animated it in code and voiced it as Earl with ElevenLabs, and Rick reviewed it on a page that pins his notes to moments in the video. The Oct 20, 2026 cut is the second: re-voiced and redrawn for the notebook design.

The prose here was written by an AI, Earl (Claude Opus 5.5), under Rick's editorial direction, from the earlier revisions Claude wrote in September. Claude, working in Claude Code on Rick's laptop, also designed and ran the experiments, built the demo and drew the figures. Rick owns the questions, the jobs, the decision to publish the losses and every editorial call.

No model scored itself, but every score rests on a model's judgment: plain code compares each answer to reference labels that Claude Fable 5.1 wrote blind, to files, before scoring. The costs are each service's own count or bill, and estimates are labeled as estimates. Jev, Gemini 3.8 Flash, Muse Spark 1.3, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 only answered.

So you're trusting the blind reference passes, the stored answers, and the rule that the design is scored once on fresh data. Each is named under Sources.

Rev 1 Sep 20, 2026: first publication.

Rev 2 Sep 20, 2026: rewritten as a guide, with the three question types first and each experiment under the point it proves.

Rev 3 Sep 20, 2026: the demo leads, a decision model reading a customer's message, code deciding and a chat model talking.

Rev 4 Sep 21, 2026: the demo's screens recaptured on the rebuilt 24-question demo. The nonstop lesson gained its repair.

Rev 5 Sep 22, 2026: rewritten around the building block, opening on the adoption figures, then the price argument, with the demo as evidence.

Rev 6 Oct 11, 2026: links to the head-to-head with OpenAI's decision model and to the build of a screen that disappears.

Rev 7 Oct 20, 2026: the explainer video at the top, narrated by Earl.

Rev 8 Oct 20, 2026: moved to the notebook design. The story comes first, every result sits beside a figure drawn for it, and each statement is marked as measured or argued. The results are numbered f1 to f15, the price argument sits on graph paper as one part, "How I built it" opens with what was written down first, and the sources are numbered by first appearance. The title changed from "The model that can't write a sentence" after a title test, and the cover is now a photograph in the house style instead of a demo screen. Every number was checked again against the run files, and four details changed. None changes a result. The 31-question tagging run cost $0.31 per thousand, not $0.29. The answer key on the attack-tool job matched the cheap model's verdict on 99.6% of rows, not 99.4%. In the demo, the decision model itself took 595 ms; 612 ms was the whole decision step. And of the ten items that fooled all three chat models, eight got the same wrong label, not all ten.

1. SourceVercel. "Jev is the fastest-adopted model in AI Gateway history," Sep 18, 2026. Quoted: "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." The metric is the share of paid AI Gateway teams; the post publishes no request or token counts. It repeats TypeSafe's benchmark, rounded to 194 and 445, and attributes it. vercel.com/blog/ai-gateway-jev-model-launchvercel-launch
2. SourceOpenRouter. Model page and provider listing, checked Sep 22, 2026: $0.042 per million input tokens, $0 per million output tokens, 32,000-token context, released on the gateway Sep 18, 2026. Usage from the public rankings feed (openrouter.ai/rankings), day view dated Sep 21, 2026: 82,379,677 requests, rank 1 of 556 models by requests and 13 of 556 by tokens; daily requests 28.2M (Sep 18), 52.6M (Sep 19), 64.4M (Sep 20), 82.4M (Sep 21). The model is served through a dedicated decisions endpoint and doesn't appear in the gateway's general model catalog, which is what cost the first morning. openrouter.ai/typesafe/jev-1.13openrouter
3. SourceTypeSafe home page, checked Sep 22, 2026: "193.6x Faster, 444.6x Cheaper," footnoted "based on workflows for System One tasks." The eval site, evals.typesafe.ai, lists four workflows (security incidents, agent-trace observability, invoice processing, customer service) and says the reference labels "are generated via an average of the responses of GPT-6 Astra and Claude Fable 5.1, both at high thinking." Graded: the vendor's own evaluation of its own model. typesafe.aitypesafe-home
4. SourceTypeSafe. "Introducing System One Models and Jev," Sep 15, 2026. On the name: "We named Jev after William Stanley Jevons." On the benchmark: the multiples are expected to be "on the higher end of real world gains," and the workflows "were made by individuals on our model capabilities team, so some bias could exist." The "System One" name is drawn from Kahneman's Thinking, Fast and Slow. The three question types are defined in the documentation at docs.typesafe.ai. Used for how the product works, not for any claim about accuracy. typesafe.ai/blog/introducing-system-one-models-and-jevtypesafe-launch
5. DerivedExperiment ledger, Sep 19, 2026: 24 numbered experiments (E0 to E24), each with its sample, seed, comparison, result, cost and per-item answer files, plus the scoring-basis correction of Sep 20, 2026 (tagging figures on the "matches either blind pass" basis: 68.2, 73.4, 74.7, 87.0). Samples of 118 to 154 items for every scored comparison; two full passes over 10,714 repositories; reference labels from nine blind Claude Fable 5.1 passes written to files. Every figure here recomputes from the raw files. Total model spend about $4.ledger
6. DerivedNorseStar Air demo: source at github.com/WorthingtonCloud/norsestar-air, live at jevdemo.worthington.cloud. Timings measured on the server as full round trips; decision-model tokens as reported by the API and priced at list; chat-model cost as billed by the gateway. First build captured Sep 20, 2026 (13 questions, 10 offers); rebuilt demo captured Sep 21, 2026 (24 questions in four parallel calls, 770 packages). Untouched captures and overlay source kept.norsestar
7. SourceThe Cherry Creek News, "Jev, TypeSafe, and the benchmark, checked," Sep 21, 2026. A secondary compilation of published real-world measurements, tallying speedups of 1.16 to 6 times, including a TypeSafe employee's launch-day test that routed one decision step of a real pipeline through the model (1.16 times faster, cost per ticket down 30.1%). Cited as one outlet's tally, not as a reproduction.cherry-creek