Reading note. This article draws on the launch of Jev on September 15, 2026, and on the coverage available at the time of writing (TechCrunch, The Register, Tom's Hardware, Simon Willison). The speed and cost figures come from TypeSafe AI itself. No technical paper has been published and no complete independent measurement exists yet. We cite them as claims, not as established facts.

In one sentence

On September 15, 2026, TypeSafe AI, a young San Francisco company founded by a former OpenAI researcher, launched Jev, an artificial intelligence model that reads text like an LLM (large language model) but never writes any: it only returns numbers, meaning choices, ratings and probabilities, each with a confidence score. The bet is simple: most AI calls inside software are not there to produce prose, but to make a call. If the bet holds, a large share of the "small model" market could shift to a new category, decision models. We think the most important signal is not the claimed speed. It is that, for the first time since ChatGPT, a serious player is betting that a more useful AI does not need to talk better.

1. A launch against the current

Let us set the scene. For three years, every major launch has pointed the same way: more reasoning, more context, more autonomy. Only three weeks ago, OpenAI presented GPT-6 Astra as the entry into "the AGI era." The race is measured in conversational and agentic capability.

TypeSafe AI heads the other way. The facts, as published:

  • The company. Founded in 2024 in San Francisco by Diogo Almeida, Erik Gafni and Sasha Sheng. Almeida spent about four years at OpenAI, where he worked on RLHF (reinforcement learning from human feedback), the technique that made ChatGPT usable.
  • The funding. A $40 million round led by the fund DCVC, at a reported valuation of around $200 million.
  • The product. Jev, available in limited early access. A proprietary model built on the Transformer architecture and trained, according to the company, exclusively on synthetic data.
  • The price. $0.042 per million input tokens, and nothing for output, according to the pricing relayed by Simon Willison and The Register.

Almeida's diagnosis, as reported by TechCrunch, fits in one idea: today's LLMs are poorly suited to automation because computers do not speak the same language as they do. Software does not want a nuanced paragraph; it wants a boolean, a category, a number.

2. What a decision model does

The change of approach deserves a precise description, because it is simpler than it looks.

With a classic LLM, you ask a question in natural language and get text back. If you want a decision a program can use, you ask the model to answer "yes" or "no," then hope it follows the instruction, adds no commentary, and does not invent a third option. A whole engineering practice has grown up to catch these deviations.

Jev removes the problem at the root. The developer describes a state (text or semi-structured data) and then asks questions in only three forms:

  • Choice — pick among options fixed in advance. Jev returns a probability distribution over those options.
  • Score — rate on a defined scale (urgency from 1 to 5, relevance, quality).
  • Noul — a closed question, which Jev answers with a probability between 0 and 1.

Two consequences follow. First, the output is constrained by design: the model cannot answer outside the schema. This is the sense in which TypeSafe AI claims to eliminate hallucinations. The claim is accurate on form, more debatable on substance: a wrong pick among three options is still an error, just a well-formatted one. Second, confidence becomes first-class data. Software can act on its own above 0.9 and call on a human below it.

RLCD — short for Reinforcement Learning for Calibrated Decisions. This is the training method the company claims. Where RLHF optimizes for what humans prefer, RLCD reportedly seeks to align stated probabilities with actual outcomes. A calibrated model is one that, when it says "80%," is right about four times out of five.

TypeSafe AI places Jev in a category it calls System One models, a reference to the fast, intuitive thinking described by psychologist Daniel Kahneman. The image is telling: frontier LLMs play the role of slow, deliberate reasoning; Jev plays the reflex.

3. The numbers, and what to take from them

The company claims responses in 70 to 500 milliseconds, and a model 40 to 200 times faster and 40 to 400 times cheaper than frontier models, with peaks of 193x and 444x measured on its own workflows. It acknowledges a "possible bias" itself, and that these peaks represent the top of the range.

Field reports are more modest, and more interesting for that reason:

  • Vercel, according to TechCrunch, reportedly replaced Luna, OpenAI's small model, with Jev on some tasks, with results 5 to 18 times faster and better accuracy.
  • Bryo AI reports a cost 10 to 20 times lower than Gemini, and above all confidence scores it can actually use.

Keep the order of magnitude rather than the record: a factor of 10 on the cost and latency of a classification task. That is more than enough to change a software architecture. At $0.042 per million tokens, a million decisions each covering a text of roughly 500 tokens would cost in the region of twenty dollars.

The limits are already documented. Simon Willison notes that the model struggles with numbers, dates and content designed to trick it. Commentator Mo Bitar asks the question that remains open: "I know it's fast. I know it's cheap, but is it good?" And above all, Jev cannot justify its decisions. An LLM can at least produce an explanation, however imperfect. Here, the reasoning disappears behind a floating-point number.

4. What it means for the market

The AI market splits into two layers

Until now, one type of tool served for everything: drafting an email, summarizing a contract, and deciding whether a message is spam. Jev formalizes a split that already existed in practice:

  • The layer that talks — writing, reasoning, conversation, agents. It remains the territory of frontier models, expensive and slow by nature.
  • The layer that decides — sorting, routing, filtering, rating, prioritizing. Billions of invisible micro-choices that keep software running.

The second layer probably accounts for most calls by volume. Until now it was billed like the first, by oversized models. That inefficiency is what Jev is exploiting.

The big labs' small models are directly exposed

Entry-level lines (the Nano, Flash and Luna versions from the big labs) are used heavily for classification. They are what Jev competes with first, not frontier models. The Vercel case should be read this way: it is not Astra being replaced, it is Luna.

We see three possible responses from the big labs: cut the prices of their small models further, expose native "decision" endpoints (an LLM's internal probabilities already exist and could be served better), or acquire. The absence of a technical paper and the reliance on synthetic data suggest that TypeSafe AI's lead rests more on execution than on a hard-to-reproduce secret. Imitators are already here: a community comparison, JevBench, reportedly lists more than 80 "Jev-class" systems. If this dynamic holds, the category could become a commodity within months.

Jevons' paradox, this time in the name

The model is named after William Stanley Jevons, the 19th-century economist who observed that making coal use more efficient increased total consumption. We described this mechanism in our piece on the collapse of token cost. Jev embraces it openly.

When a decision costs a fraction of a cent and a tenth of a second, you do not simply replace existing calls: you add them everywhere you never would have dared. Every row of a spreadsheet, every event in a technical log, every action of an agent can get its score. Unit spending collapses; volume, on the other hand, could explode.

"Smarter systems, not smarter models"

The phrase belongs to Adam Jacob, as relayed by The Register, and it sums up the shift in value. Andrej Karpathy sees it as an answer to latent demand, under-invested because the whole industry was chasing more intelligence. If this reading is right, competitive advantage moves from the model to orchestration: knowing which question to send to which model, at what confidence threshold to escalate, when to bring in a human.

One use case illustrates this shift well: agent monitoring. An expensive autonomous agent can be supervised by a decision model that checks, at every step, whether the planned action stays within bounds. The reflex watches over the reasoning.

Evaluation and compliance become the real product

Jev's opacity has a direct consequence: it can only be trusted if it is measured. The paradox is that its negligible cost is exactly what makes exhaustive evaluation possible, across tens of thousands of cases. We expect to see a market emerge for evaluation and calibration tools specialized for these models.

On the regulatory side, the temptation will be strong to use a decision model to screen job applications, score credit files or prioritize requests for assistance. Yet the European Union's AI regulation (the AI Act) already classifies several of these uses, such as recruitment, as high-risk systems, with obligations of transparency and human oversight. A score without an explanation is a poor fit. Simon Willison also flags possible biases in his very first tests. In Europe, the pace of adoption of decision models may depend less on their performance than on the ability to audit their choices.

5. Signals to watch

  1. Independent measurements. As long as the 193x and 444x figures are measured by TypeSafe AI on its own workflows, they remain sales arguments. A third-party benchmark on public tasks will show where reality lies.
  2. Actual calibration. The whole promise rests on reliable probabilities. If a Jev "0.9" proves right nine times out of ten, across varied domains, the category is solid. If not, it is just a fast classifier.
  3. The big labs' response. A native decision API from OpenAI, Anthropic or Google, or another price cut on their small models, would confirm that the threat is being taken seriously.
  4. Equivalent open models. If an open-weight decision model appears, the category will escape the control of a single player and become a standard building block.
  5. The first sensitive uses. Recruitment, credit, moderation: this is where the first controversies, and the first demands for explainability, will surface.

6. A situated word

From Réunion Island, Jev speaks to us for a simple reason: it vindicates an intuition we have defended since the lab began. The right answer is not always the most powerful model. Much of what we ask of AI in a concrete tool, whether sorting messages, spotting an emergency or routing a request, is a matter of reflex, not reflection. Paying a frontier model for these tasks is like chartering a plane to cross the street.

For a frugal, island-based player, three lessons stand out.

First, the cost of a decision is no longer a barrier. A small organization can now add automatic judgment to its tools for a few dollars a month. The argument "AI is too expensive for us" loses much of its force.

Second, network latency matters more than model latency. When the model answers in 150 milliseconds, the round trip between our island and a Californian server can weigh as much as the computation itself. The faster the model, the more visible the distance. That is one more argument for keeping a modest capability locally, in the spirit of what we wrote about desktop hardware able to run models.

Finally, a rented decision remains a revocable decision. Jev is proprietary, hosted in the United States, with no technical publication. We saw it with Fable 5: access can vanish within days, and demand at launch was so strong that the API reportedly went briefly unavailable. The right approach, in our view, is this: make use of the tool, but design our systems so the decision layer stays replaceable, with our own test sets to check that another model does just as well.

There is something reassuring about this launch. After three years of a race in which every announcement promised a more general intelligence, someone is reminding us that useful software is often made of small decisions that are right, fast and cheap. It is not the most spectacular direction, but it may be the one that reaches the most people. 決

Sources and further reading

This document is updated if new elements emerge. Last revision: 文 September 24, 2026.