TypeSafe AI launches Jev, a viral new 'decision model' category priced per input token only
2026-09-21
TypeSafe AI launched Jev, described by both Latent Space and Simon Willison on Sept 21, 2026 as a new category of 'System One' or 'decision' model: it accepts text/structured input but returns only floating-point probabilities, confidence-scored choices, or numeric scores rather than generated text. Willison reports Jev's input price is $0.042 per million tokens with no charge for output, and that TypeSafe's own comparison puts its launch video at roughly 40 million views, more than GPT-4o's (22M) or Claude Fable 5's (15M) launch videos. Willison tested Jev by scoring San Francisco Bay Area cities on a yes/no 'Good city?' question, found it rated Cupertino highest and East Palo Alto lowest, and cautioned that the lack of explainability makes hidden bias hard to detect, specifically saying he hopes nobody uses it to rank job applicants.
Significance 4: A genuinely new model category from a new entrant, independently described and tested (not just relayed) by two separate curators, with viral-scale distribution comparable to a major lab's flagship launch — treated as a significant new entrant per the rubric, though unproven at production scale.
Implications · machine-drafted, not owner judgment
Jev's economics (input-only pricing, sub-cent per million tokens) make classification, ranking, and reranking dramatically cheaper than routing them through a full LLM, which could shift how agentic pipelines are built around cheap decision layers. Willison's bias concern — opaque floating-point scores with no rationale — is a concrete instance of the black-box risk the owner's worldview flags for any tool used to judge people, and a sharp real-world argument for why this kind of model should not be used for consequential decisions like hiring without independent bias auditing.
- Whether Jev or a similar decision-model product gets adopted inside a major agent framework or coding tool within the next month or two
- Any documented bias audit or independent evaluation of Jev's scoring behavior
Sources
- curator Jev: System One models for Prod, not God retrieved 2026-09-22
- curator Jev introduces a new shape of LLM - System One, aka Decision Models retrieved 2026-09-22