Jev

Most models on this hub answer in sentences. Jev answers in types.

TypeSafe opened early access to it on 15 September 2026 and calls it the first System One model: you give it a block of state and a set of questions, each declared as a choice, a true/false or a score, and it returns an answer to every one of them in a single call, with a probability attached. The company’s founder is Diogo Almeida, who worked at OpenAI on the methods behind ChatGPT.

You hand it state and typed questions

A call carries three things: the model name, a state string holding whatever the software knows, and a questions object. Each question declares its own shape. A choice names its options, a noul gives the true and false conditions, and a score lists ordered anchors.

YFarmX sent it a support email and asked three questions at once. This is the answer that came back, trimmed to the fields that carry the result:

{
  "model": "typesafe/jev-1.13-20260917",
  "answers": {
    "is_support_request": { "type": "noul", "noul": 0.97 },
    "route_to": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 1, "technical": 0, "account": 0, "sales": 0 },
      "confidence": 1
    },
    "urgency": {
      "type": "score",
      "score": 2.32,
      "legend": { "0": "Routine", "1": "Elevated", "2": "High", "3": "Critical" },
      "probabilities": { "0": 0, "1": 0.01, "2": 0.66, "3": 0.33 },
      "confidence": 0.66
    }
  },
  "usage": { "input_tokens": 509, "output_tokens": 79, "cost": 0.000021378 }
}

The useful part is the third answer. The email described a card declined twice with a suspension due on Friday, and the model put 0.66 on “High” and 0.33 on “Critical” rather than picking one and discarding the doubt. The score, 2.32, is where those weights land on the four-point scale.

We timed it on a support inbox

Thirty-one identical calls, each carrying the three questions above, run from a container on 18 September 2026 against ~typesafe/jev-latest:

Most calls landed between 185 ms and 320 ms. Two sat outside that: the first call of the run at 433 ms, which is the shape a cold connection usually makes, and one later call at 982 ms.

The cost is worth checking against the published price, because it confirms it exactly: 509 tokens at $0.042 per million is $0.000021378, which is the figure the response itself reported. Output tokens were billed at zero. At that rate a million routing decisions of this size cost about $21.

The score moves while the decision holds

Running the same call fourteen times showed something a builder needs to know before wiring Jev into a test.

The decision was stable. route_to came back billing all fourteen times, at probability 1.0. The boolean came back 0.97 every single time.

The score moved: 2.26, 2.27, 2.29, 2.31, 2.32, 2.33, six distinct values across a range of 0.07. It stayed inside the “High” band throughout, so every run agreed on the answer while disagreeing in the second decimal place.

This has a practical edge. A score is continuous across the bands, so code that compares it to an exact value, or looks it up as an integer, finds nothing. Our own model-guard command line tool hit exactly this: both of its score rows printed “unavailable” on every live run until they were rounded onto the nearest band. Read the band, not the decimal.

What people have built with it

Early access opened on 15 September 2026, and by 18 September the community list at awesome-jev was carrying more than a hundred public projects. The table below holds the ones YFarmX opened and confirmed, ordered by GitHub stars read on 18 September 2026.

Checking them one by one was worth the time. Three of that list’s best-known entries describe no Jev use at all on their own pages, including a project with 23,600 stars, so every row below was read on the project’s own repository before it went in.

Two of those carry measurements worth repeating.

Ably Labs timed Jev against chat models on the same task, recorded on 17 September 2026: Jev averaged 227 ms and 4.4 decisions a second, against 2.5 s for Claude Haiku 4.5, 3.2 s for Gemini 3.8 Flash and 3.5 s for GPT-5.6 Sol. Its own note records that the chat models answered correctly 95 to 100% of the time, so what the demo shows is the speed rather than the judgment.

The Pokémon agent published its running costs: about one decision a second, 1,600 to 2,100 input tokens per decision, and roughly $0.27 an hour, or $0.76 per 10,000 decisions. It reaches the starter Pokémon and returns to Pallet Town without help.

Where it runs and what it costs

Jev is served on OpenRouter’s alpha decisions route, POST /api/alpha/decisions, billed to an OpenRouter account, and through the Vercel AI Gateway. Input is $0.042 per million tokens and output is free. Context is 32,000 tokens with a 28,800 token output ceiling, and a single choice question can carry up to 255 options.

Its first days carried real volume. The provider page recorded 56.5 billion tokens processed on Jev 1.13 by 18 September 2026, three days after early access opened.

One thing to expect when looking it up: the model is absent from OpenRouter’s /api/v1/models catalogue, because it answers on the decisions route rather than chat completions. A catalogue lookup returns nothing and the endpoint still answers. Probe the endpoint itself.

The build served to us on 18 September identified itself as typesafe/jev-1.13-20260917, a dated build of the 1.13 line, and the provider field read TypeSafe.

How the fingerprint reads

YFarmX measures the tokenizer signature of every model in its catalogue, and Jev has a record like any other. It was measured on 18 September 2026 across the decisions route: 50 clean rows, no corrupt rows, signature tk_5e398719.

Its declared tokenizer family is “Other”. The closest model in the catalogue is Tencent’s Hunyuan A13B Instruct, which agrees on 24 of 50 probes, with Tencent’s HY-MT2-7B level at 24 and Poolside’s Laguna XS 2.1 at 21. Agreement at that level is what two unrelated tokenizers look like, so the signature reads as its own: distinct from everything measured so far, and held open for the day a closer match arrives.

Read the record and compare it against anything else in the catalogue:

• Jev 1.13’s own catalogue record
• Jev 1.13 against Hunyuan A13B Instruct, its nearest neighbour
• The fingerprint comparison tool, for any pair in the catalogue
• How the fingerprinting works, and what the probes measure