TypeSafe's Jev Wants to Take the LLM Out of the Decision Loop
TypeSafe released Jev yesterday. The speed numbers got the attention. What stayed with me is that they trained a model to do only the deciding part. New architecture, parallel sampling instead of one token at a time, and a training method aimed at calibrated decisions rather than human preference. Unstructured state in, typed values with probabilities out, in about 100ms.
Today we build agents by wrapping a loop around a text generator. The model writes tokens, your code parses them, and the decisions that drive the whole system come out of the same machinery that writes prose. That shape is an accident of what was available.
Dropping language generation entirely is where the efficiency comes from. Two orders of magnitude more intelligence per euro and per millisecond, because none of the capacity goes into sounding fluent.
And once deciding costs almost nothing, deciding can own the loop.
An action model reads the state, picks the next action, routes, scores, verifies, and repeats that hundreds of times a second. Text generation becomes one of the actions available to it. You call an LLM when the output genuinely has to be language: a reply to a customer, a diff, a summary. The rest of the time control flow never leaves the decision layer.
The large general model stops sitting at the centre and becomes a tool, and the thing holding the loop together is small, typed and calibrated.
The part I keep coming back to is that Jev is their first model, and a deliberately cheap one. It already sits at frontier level on this class of task at a fraction of the cost per call. Nobody has spent LLM-scale money on this architecture yet. When someone does, the budget we hand an LLM today should buy a meaningfully smarter decision than anything we can buy now.
I am not calling any of this settled, and the evals so far are TypeSafe’s own. But this is the first time I have seen the architecture, the training method and the price line up in the same direction.
This one stands out — it's about a genuinely brand-new product (launched September 15, 2026, so "yesterday" checks out exactly), and unusually, the post's own epistemic hedging matches almost exactly what independent tech reporters are flagging too. That's rare and worth acknowledging directly.
TypeSafe and Jev are real. Founded by Diogo Almeida (credited as a co-inventor of ChatGPT at OpenAI) along with Erik Gafni and Sasha Sheng, TypeSafe raised roughly $40 million in seed funding led by DCVC and released Jev in early access on September 15, 2026.
The architecture description is accurate, not simplified into inaccuracy. Jev uses a parallel sampler that returns typed, probabilistic decisions in a single pass rather than generating tokens autoregressively, and it's trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), explicitly positioned against RLHF/RLVR — optimizing for calibrated confidence on structured decisions rather than human-preferred text. Matches your description precisely.
"~100ms, unstructured state in, typed values with probabilities out" — accurate. TypeSafe reports end-to-end latency of 70-500ms, and their own tagline is literally: unstructured state in, typed probabilistic decisions out. "About 100ms" sits on the fast end of that range but isn't fabricated.
Pricing and speed multiples — confirmed. $0.042 per million input tokens with free output (versus $2-12/MTok for comparable frontier LLMs), and TypeSafe's own workflow benchmarks report speed gains from 40x up to a headline 193.6x on favorable tasks. Your post sticks to the more conservative "two orders of magnitude" framing rather than reaching for the 193x/444x headline numbers — which is the right call, since multiple outlets covering the launch explicitly note TypeSafe itself says those bigger figures sit near the high end of what customers should expect, not the typical result.
Your own caveat — "the evals so far are TypeSafe's own" — is exactly right, and matches independent scrutiny almost point for point. Reporters covering this launch note the release describes the approach at a high level rather than publishing model weights or a paper with enough detail for outside researchers to reproduce the results, and list open questions that are essentially unanswered right now: how the architecture really differs from a transformer beyond "parallel sampler," whether calibration holds under distribution shift or adversarial inputs, and critically, how much of the cost advantage survives once a larger model has to step in to explain or verify a low-confidence decision. That last point is worth adding to your post specifically, since it's the most direct challenge to the "two orders of magnitude cheaper" framing in real deployment.
One editorializing note: the "large general model becomes a tool, deciding never leaves the decision layer" framing is your own forward-looking thesis, not a claim TypeSafe has made about the industry generally — reasonable and well-grounded in Jev's actual design (developers keep code in control of the workflow; Jev supplies judgments), but worth knowing it's your extrapolation, not their marketing claim.