TypeSafe AI, led by former OpenAI researcher Diogo Almeida, released a new model class called System One Models on September 15 local time, along with its first model, Jev. The design is deliberately narrow: Jev never generates text, and instead returns only decisions that follow a schema defined in advance. End-to-end responses land between 0.07 and 0.5 seconds. Early access signups opened the same day.

Not an Extension of Chat, but Decisions Used Inside Software

Almeida says the question driving him for the past 4 years was why automation had not arrived, given that models had been superhuman at conversation for years. At OpenAI he worked on the methods that taught models to follow instructions and hold a conversation with people, research that became the foundation of ChatGPT. Even so, he stopped believing that AGI lay somewhere past chat, and spent 2 years in stealth digging in a different direction.

The result is a new model category called System One Models. The name comes from Daniel Kahneman's Thinking, Fast and Slow and its split between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning. Where today's LLMs lean toward the System 2 role, this class commits entirely to fast, structured judgment.

Jev itself is named after the economist William Stanley Jevons. The reference points to the paradox in which improved steam engine efficiency increased demand for coal, reflecting the view that every order-of-magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases.

Giving Up Text in Exchange for Type Safety and Confidence

The sharpest decision in Jev's design is that it does not output strings. A typical LLM generates tokens one at a time, and that accumulation becomes a chat reply or a block of code. Jev emits the available options and their probabilities in parallel, in a single pass, following a schema set beforehand. TypeSafe AI describes it as a function call where unstructured state goes in and typed probabilistic decisions come out.

With that shape, output falling outside the schema becomes impossible in principle. A tool call containing a function name that does not exist, or a sentence landing in a field that should hold a number, simply stops happening. Every decision also carries a calibrated confidence value. The company's framing is that a model able to solve a task 95 percent of the time cannot be used for automation if it never flags the remaining 5 percent.

The training method was purpose-built as well. Rather than RLHF, which rewards human preference, or RLVR, which rewards verifiable answers, Jev uses RLCD (Reinforcement Learning for Calibrated Decisions), which rewards honest probabilities. Intended uses include classification, routing, extraction, and scoring, the places where hand-written if-statements turn brittle. Scoring the output of an LLM or acting as a guardrail around it are also listed.

0.042 USD per Million Input Tokens, with Output Free

Pricing sits on a different scale from existing models. Jev costs 0.042 USD (about 6 yen) per million input tokens, and output tokens are free. TypeSafe AI says it lists output at zero because the tokens are too cheap to meter. By the company's comparison, input for existing LLMs runs from 0.20 USD (about 31 yen) to 10 USD (about 1,550 yen) per million tokens, with output roughly 5 times the price of input.

※1 USD = 155 JPY

The basis for the low price is the change in how output is produced. Dropping sequential token generation in favor of a parallel pass keeps the hardware busy rather than idle. The company frames this substitution as the same kind of shift that let Transformers replace RNNs.

Responses in 0.07 to 0.5 Seconds, Fast Enough to Play DOOM

The speed gap is more extreme still. According to TypeSafe AI, frontier LLMs take 3 to 329 seconds end to end, while Jev returns in 0.07 to 0.5 seconds. The former is fine for talking with a person, but inside code that wait becomes the bottleneck.

As a legible demonstration, the company published video of Jev playing the classic game DOOM. It issues roughly 10 queries per second, and an hour of play cost about 7 USD (about 1,090 yen). The engineer behind it had worried about querying 10 times a second, though the rest of the team read the figure as cheaper than expected. The material Jev reasons over is structured game state data, not the rendered screen.

The second demo is a race from one Wikipedia article to another using only links encountered along the way. Each step can mean choosing among hundreds or thousands of links, so the benefit of never inventing a link that does not exist matters as much as raw speed. The company notes that every model ran in fast, non-reasoning mode for this demo, which keeps the speed gap modest.

How to Read the 193x Figure

The 193.6x faster and 444.6x cheaper figures on the company's home page come from a new workflow evaluation it built. Each model receives the same workflow, and results are compared against reference probabilities taken from the predictions of the largest and most expensive external models. The reference is the average of GPT-6 Astra and Fable 5.1, and TypeSafe AI acknowledges on its own that this choice biases results toward the models from OpenAI and Anthropic.

That candor runs through the whole post. The workflows were built by the company's own model capabilities team and may carry bias, the speed measurements are run from laptops on the US West Coast, and the sustainability of the pricing can only be proven over the long term. The 0 percent type error rate is not measured either: schema matching is guaranteed, so the company states it can confidently plot zero.

Read the other way, every advantage shown here rests on the company's own evaluation, with no third-party reproduction yet. Details can be traced on the published evaluation site, but 193x is better understood as a near-upper bound under favorable conditions than as a number that will appear in everyday production work.

Summary

Jev, released by TypeSafe AI, abandons text generation and returns only structured decisions, committing to a direction separate from the LLMs built so far. The headline traits are 0.042 USD per million input tokens with free output, responses in 0.07 to 0.5 seconds, and output that cannot fall outside the schema. The 193.6x speed figure rests on in-house evaluation, but the underlying idea of designing a model for use inside code rather than inside a chat window looks likely to spread to other camps. Early access registration is open, and the company is working through its waitlist.