While chatbots dominate the headlines, a company building an AI model that “cannot chat” has become a magnet for capital and enterprise adoption. TypeSafe AI, the developer behind Jev, recently closed an 870 mi l l i o n f u n d in g r o u n d a t a 7.5 billion valuation. Andreessen Horowitz led the round, with Sequoia and existing investor DCVC also participating—just weeks after Jev’s launch.
Jev is not a large language model. It does not output text; it outputs probabilities—or, in TypeSafe’s words, “calibrated decisions.” Traditional LLMs produce natural language, after which developers must write extra parsing code to extract the needed value. Jev instead maps inputs into a preset output set and directly returns yes/no judgments, numerical scores, or fixed-option results—typed values that code can consume immediately. TypeSafe calls this approach “reinforcement learning for calibrated decision-making.” Co-founder Diogo Almeida, formerly at OpenAI and a co-inventor of RLHF, left OpenAI because he observed that RLHF makes models better at conversation while quietly degrading their calibration, leading to “mode collapse.”
On the performance side, Jev responds in under 700 milliseconds, roughly 100× faster than frontier LLMs and at 1/100 to 1/500 of the cost. It is priced at 0.042 p er mi l l i o nin p u tt o k e n s , w i t h o u tp u tt o k e n s f r ee . I nan o f f i c ia l co m p a r i so n , a co mm e n t − ann o t a t i o n t a s k t oo k J e v 203.2 seco n d s an d 0.84, versus 823.5 seconds and $1.50 for a competing LLM. Within three days of launch, Jev had generated one trillion tokens.
Enterprise uptake has been equally rapid: TypeSafe claims one-third of Fortune 500 companies are already using Jev for spam filtering, trading decision support, agent action adjudication, workflow approval, classification, and routing. The buzz has also triggered fast-follow moves across Silicon Valley: OpenAI introduced the Decisions API, Databricks launched ai_decide, Cloudflare open-sourced the Clef family, and Amazon already fields Strands Decider 2B. The emerging division of labor is clear—generative models handle complex reasoning, while decision models handle high-certainty automation.
Top comments (0)