Three Weeks Old and Already at a Trillion Tokens a Day: Jev and the Rush to Clone the Decision Model

According to a Wall Street Journal report published Monday, Jev — the decision model that launched just three weeks ago — is now processing roughly 1 trillion tokens a day, and about a quarter of the Fortune Global 500 have put it to work. Both figures come from TypeSafe AI itself and have not been independently audited, but they describe an adoption curve few enterprise software products have ever traced.

Jev does something deliberately unglamorous. Instead of generating free-form text, it takes an input and assigns it to a preset output: yes or no, a numeric score, or one option from a fixed list. TypeSafe calls the technique reinforcement learning for calibrated decisions. The pitch is that most of what software needs from AI day to day is not eloquence but repeatable judgment — route this ticket, approve this refund, block this agent action — and that a model built to answer bounded questions can be faster, cheaper and more reproducible than a chat model.

The company was founded by Diogo Almeida, a former OpenAI engineer who worked on ChatGPT, and has raised $40 million from DCVC. The Journal's report arrives amid separate reports — first from The Information — that TypeSafe is in talks for a new funding round at valuations above $10 billion. Almeida declined to comment on the fundraising.

The more telling part of the story is who is copying it. OpenAI shipped a Decisions API built on its Luna model that answers user-defined questions within a fixed set of outcomes. Databricks followed a day later with ai_decide, a sub-second structured decision function. Cloudflare open-sourced Clef and Clef-flash, built on frozen Qwen3.8-27B and Qwen3.5-9B bases with a dedicated routing head that scores candidate paths without autoregressive generation. Amazon already had Strands Decider 2B in its Strands Agents SDK.

A category is consolidating in real time: generative frontier models for open-ended reasoning, and a separate class of small, cheap, calibrated classifiers for the high-frequency, low-ambiguity decisions underneath agent systems — routing, classification, approvals, and gating what an autonomous agent is allowed to do next. The division of labor echoes what happened in infrastructure years ago, when specialized hardware took over workloads general-purpose CPUs handled badly.

Almeida told the Journal that competitors were inevitable and framed Jev as the opening move in a new class of AI rather than a product with a moat. There are still many frontiers left to explore, he argued; TypeSafe took the first step into the next one, and the industry will keep finding directions beyond it.

The caveats are the usual ones for a hype cycle this fast: the token and customer figures are self-reported, the round is unconfirmed, and three weeks is too short to know how these models hold up in production. But the direction is no longer in question. Every major platform vendor now ships a decision-model product, which is about as strong a signal as a nascent category gets.