TypeSafe AI says Jev is 193.6x faster than Claude Sonnet 5

On September 15, 2026, a company called TypeSafe AI announced a new model named 'Jev.' The development is led by Diogo Almeida, a co-author of the paper (InstructGPT) that served as the foundation for ChatGPT. After about two years in stealth mode, this model emerged with $40 million in seed funding, and its most prominent feature is that it 'does not generate text at all.'

A commitment to returning only choices and probabilities

What Jev can do is return only the selected value, its probability, and a confidence score when provided with predetermined choices and data. It cannot reply with text, generate code, summarize, or explain its reasoning. It specializes in 'judgments for code to branch subsequent processing,' such as classification, sorting, and scoring. Because it is designed to select answers only from a fixed set of options, the selling point is that, in principle, hallucinations (plausible lies) cannot occur. Processing is claimed to be 70–500 milliseconds, at a cost of $0.042 per 1 million input tokens.

The design philosophy visible in the name 'System One'

Personally, what I found most interesting is that this model carries the category name 'System One.' This is a reference to psychologist Daniel Kahneman's distinction in 'Thinking, Fast and Slow' between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2). TypeSafe AI positions current LLMs as leaning toward System 2 because they possess chain-of-thought, which accumulates the process of thinking. It is interesting to see one of the key figures behind ChatGPT move away from the 'make one model do everything' design and steer toward a division of labor where 'fast judgments are handled by Jev, and deliberate generation is handled by existing LLMs,' as it serves as one answer to how AI should be built.

It is better to view the numbers with a discount on the 'maximum values'

TypeSafe AI itself touts figures such as '193.6x faster than Claude Sonnet 5 and 444.6x cheaper than Opus 5,' but these are maximum values for specific tasks involving multi-stage judgments. Independent verification by a third party (Every) yielded results that, while similar in direction, differ in scale, showing it to be 'about 25x faster and 580x cheaper than Claude Fable 5.1' in extraction tasks. The demo that became a hot topic as a demonstration—'having Jev play DOOM in real-time'—was also intended to prove a reaction speed of about 10 judgments per second; the skill of the gameplay itself is not the main point (it is said that a simple scripted bot would be stronger). It seems closer to reality to look at the depth of such substantiation rather than flashy multiplier numbers.

Summary

Jev is not a replacement for conversational AI, but rather a dedicated component that isolates only the 'judgment' process. I want to keep an eye on whether this design philosophy—dividing roles between a high-speed judgment layer and a slow-thinking generation layer, rather than making one generative AI do everything—will spread in the future.