New AI Models That Decide Instead of Talk: What Jev Proves and What It Doesn't

The category arrived in two weeks

The new AI models getting attention this fall do not write anything. TypeSafe AI, a San Francisco lab that came out of stealth with a $40M seed round led by DCVC, launched Jev on September 15, 2026. You send it a piece of state and a set of typed questions, and it returns a choice, a score or a probability instead of text. Input costs $0.042 per million tokens, output is free, and TypeSafe quotes 70 to 500 milliseconds end to end. TypeSafe calls the category System One models.

Then the market moved fast. OpenAI announced a Decisions API at DevDay on September 29, built on a specialized version of GPT-6 Luna. As of early October it was still in limited preview with no published pricing. On October 1, Cloudflare released open-weight Clef and Clef-flash models under Apache 2.0, and AWS released Strands Decider 2B. Vercel said Jev reached nearly 13% of its paid teams within 24 hours of landing on AI Gateway, twice the share of the GPT-5.6 family.

For US teams, the speed of the copying is the signal. Three major American AI and cloud companies shipped their own version of the idea within two weeks of a startup's launch. That says the decision layer is a real product category. It does not yet say TypeSafe's approach is the one that wins.

A new interface, an unproven architecture

The headline claim is that Jev is a different kind of model from a traditional LLM. The evidence supports a narrower version. TypeSafe says Jev uses a new architecture, a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions. It has not published a paper, and a review of the independent evidence found no disclosed model size, base model or training data.

Everything that has shipped since is built on existing language models. Cloudflare trained Clef from Qwen3.8-27B and Clef-flash from Qwen3.5-9B. AWS started Strands Decider 2B from Qwen3.5-2B and replaced the text output with a pointer head of about 1 million parameters. An open-weight student of Jev on Hugging Face runs on Qwen3.5-9B.

Researchers at the University of Bonn made the same point from the other direction. They scored Qwen3.8-27B and Gemma-4-E4B as decision models by reading each model's next-token probabilities over the answer options, using identical requests. Jev beat Qwen on 27 of 37 datasets and Gemma on all 37. So the decision interface is not unique to Jev, but Jev's training appears to add something on top of it.

That makes "a new kind of AI model" too strong. What is new is a product category: models that return typed, calibrated decisions and are priced like infrastructure. Decision models are not simply small language models either. They reuse language-model backbones for a different job, picking from a fixed list instead of writing. The types of AI models a team chooses from now include a decision layer that sits next to the generative one, not in place of it.

What the independent AI model comparison shows

The broadest independent test so far comes from the University of Bonn. The researchers ran Jev on 37 public datasets, 346,009 requests in all, for under $10. Jev reached 95% to 99% accuracy on IMDB, SST-2, HellaSwag and ARC, and 86.7% on Belebele across 122 languages. It struggled on low-resource languages, noisy or fine-grained labels and rubric-style quality judgments. The authors report that its choice probabilities are well calibrated.

A community benchmark of 868 real decisions from the n8n repository, labeled mechanically from what each code change actually did, put Jev close to the frontier but behind it. On a routing task Jev scored 85.6%, against 88.3% for GPT-5.6 Terra and 90.6% for Claude Opus 5. On four-way triage it scored 70.9% against 73.0% and 80.4%. On a risk question it scored 63.9% against 66.1% and 80.0%. The benchmark's author says it is not neutral, and published a correction after an earlier headline result failed to replicate.

The wider record is consistent. In a pre-registered social-science annotation study with 7,977 human-labeled items, Jev trailed the best of 19 LLMs on 14 of 15 tasks by a median of 11.6 macro-F1 points, according to a DEV Community review that re-scored the published outputs. Where the job is screening rather than judgment, Jev holds up: an arXiv study of 5,219 agent trajectories found a Jev-based judge averaged an F1 of 77.8, against 74.1 for the strongest generative judge.

Taken together, the independent data puts Jev in the middle of the pack on accuracy. It is strong on binary and few-option decisions about the gist of a short text, level with mid-priced LLMs, and several points behind the frontier on harder calls. That is a useful tier, not a replacement for the best models. One caveat: most of these studies are days-old preprints or community repositories, and none has been through peer review.

The speed and cost claims shrink in the wild

TypeSafe's headline figures are 193.6 times faster and 444.6 times cheaper than frontier LLMs. Its own launch post calls those the high end of real-world gains and discloses that its model team built the test workflows, that accuracy is scored against the average of GPT-6 Astra and Claude Fable 5.1, and that the LLM baselines ran through TypeSafe's own adapter. It also says it cannot prove its pricing isn't subsidized.

Independent measurements land lower on speed and still high on cost. The n8n benchmark measured Jev at 5.7 to 11.8 times faster and 120 to 242 times cheaper per decision than Claude Opus 5, with median latency around 410 to 430 milliseconds. An analysis of 12,759 launch posts found a median user-reported speedup of 7x, a median cost saving of 30x and a median latency of 76 milliseconds. Across the independent studies the DEV review traced, measured speed gains ran from 0.5x to 12.1x and cost gains from 0.6x to 478x, depending on the comparison model.

Most of the cost gap comes from output being free. In one workload in that review, 83% of the old bill was output tokens. The review's author also measured a plain open model, Gemma 4 26B, on one EC2 L4 GPU at full load at $5.43 per million decisions at most, against $5.54 for Jev at the same prompt size. Low cost per decision does not have to depend on one startup's launch pricing.

TypeSafe named Jev after the Jevons paradox, the idea that cheaper intelligence unlocks more uses. If that logic holds, cheap decisions will show up inside far more software than chat ever reached. AETW has covered the same dynamic in the labor data.

The catch: a valid answer is not a correct answer

Type safety means Jev cannot return something outside the options you defined. It does not mean the answer is right. InfoQ's coverage captured the objection from a Hacker News commenter: the model cannot emit an invalid type, but it can emit a wrong valid value. Armin Ronacher, CTO of Earendil, told TechCrunch the design "delegates the hallucination problem a little bit to the user," who must decide whether a 50% probability is worth acting on.

Calibration is the real test, because the pitch is that you act above a threshold and escalate below it. The Bonn team found calibrated choice probabilities, but also that the default 0.5 cutoff on yes-or-no questions often turns good ranking into mediocre decisions. On prompt-injection detection, precision was perfect but recall was only 50%, even though ranking quality (AUROC) was 0.982. Thresholds tuned on training data lifted micro-F1 on one legal-clause set from 0.50 to 0.75. In the n8n benchmark, Jev claimed 0.9 to 1.0 confidence on 75 size decisions and was right 48% of the time. The DEV review found the direction of the calibration error changes by domain, and that fitting a single temperature on 50 to a few hundred labeled examples fixes most of it.

TypeSafe's own jaggedness page, last reviewed October 2, 2026, lists the known failure modes: literal reading, counting and arithmetic, date comparison, large inputs full of irrelevant detail, adversarial content and sensitivity to option order. Its advice is to keep math and dates in code. The same DEV review reports that models trained on a task's own labels beat Jev wherever they were tried, including a 310M-parameter encoder trained on 200 rows that beat it by 12 points on Japanese news topics.

What US teams should do with this

• Use decision models for routing, classification, screening and scoring where a wrong answer is cheap to catch. Keep generation, long reasoning, arithmetic and date logic in an LLM or in code.
• Benchmark on your own data before switching. The n8n benchmark's author found a 29.8-point gap to Claude Opus 5 on a triage task in a private repo, then 9.5 points on a public one.
• Label 50 to a few hundred examples and set your own thresholds instead of trusting the default 0.5 cutoff.
• Pin an exact model version such as jev-1.13.0 rather than the moving latest alias.
• Test the open alternatives. Cloudflare's Clef and AWS's Strands Decider 2B ship under Apache 2.0, and AWS frames its model as part of hybrid agents where a decision model picks and an LLM reasons.
• Treat launch pricing as unproven. TypeSafe says it can't prove its prices aren't subsidized, and OpenAI has not published a Decisions API price.
• Watch OpenAI's broad release of the Decisions API, which should bring the first pricing and accuracy numbers.

Sources