Awesome Search KG

TypeSafe

6 min read

TypeSafe

TypeSafe is an AI company building what it calls AI primitives — modular, composable decision components intended to be reliable and fast enough for software to consume directly. Its flagship model is Jev, which the company positions as the first System One model: a model that makes fast, structured decisions and returns typed values and probability distributions for code, rather than text for a person.

The framing is a deliberate contrast with general chat models. A System One model generates no free-form text, no code, and no written rationale; the possible outputs are fixed by the caller’s schema before the call is made.

Website: https://typesafe.ai · Documentation: https://docs.typesafe.ai/introduction

Company

Founded by Diogo Almeida, previously at OpenAI, where he describes having helped build the instruction-following and conversational methods behind ChatGPT. TypeSafe spent roughly two years in stealth before announcing Jev and the System One Model class in September 2026 (Introducing System One Models & Jev).

The founding premise is that chat-shaped models are the wrong interface for software: strings must be parsed and validated, and a model free to emit anything can always go off the rails. The company’s bet is the inverse — a fixed output space, parallel sampling, and training for calibrated confidence (Reinforcement Learning for Calibrated Decisions) rather than human preference.

Both names come from the FAQ in that announcement: System One from Daniel Kahneman’s fast, intuitive System 1 thinking, and Jev from the economist William Stanley Jevons — the argument being that each order-of-magnitude drop in the cost of intelligence unlocks orders of magnitude more uses.

Jev launched on 2026-09-15 alongside a reported $40M round led by DCVC, per How to Use Jev - A Practical Guide.

Reading the Company’s Claims

TypeSafe’s headline figures (193.6× faster, 444.6× cheaper) come from workflow evals, an evaluation format it designed: a fixed compute graph given to every model, scored not against ground-truth labels but against the averaged predictions of large external models as reference probabilities. The announcement is unusually explicit about the resulting biases — the workflows were built by its own model-capabilities team, the reference models bias the comparison toward those vendors, the competing LLMs run through TypeSafe’s own wrapper, and its own 0% type-error figure is derived by construction rather than measured.

That candour is worth crediting, but the structure stands: the architecture claims are checkable, while the intelligence-parity claim rests on a benchmark the vendor designed and scored.

TypeSafe is not a search company and does not market Jev as a reranker. It appears in this vault because the model’s call shape — a shared state plus many independent typed questions answered in parallel with probabilities — turns out to fit Reranking closely enough to compete with purpose-built rerankers without any reranker training.

Hev meets Jev is the benchmark that makes this case: on three BEIR subsets, an untuned Noul-per-document prompt landed at 0.501 mean nDCG@10, against 0.504 for Voyage rerank-3 and 0.486 for Cohere rerank-v3.5. The conclusion drawn there is less about TypeSafe than about the category: a general decision model with no ranking training now sits in the same quality and price bracket as the reranking specialists, and adds a calibrated probability none of them return.

Two caveats recorded in the same source: Jev is hosted only, so documents leave the caller’s environment, and its 32k-token request budget bounds rerank depth at roughly 50 passages.

How the Pricing Claim Reads Under Audit

The launch’s sharpest number is a 238x lower input price than Claude Fable 5.1. Sajith K checked it — 0.042 per million input tokens — and found the arithmetic exactly right, while pointing out that the comparator is a frontier model at the top of the price curve rather than the model anyone would actually use for classification work. Against Claude Haiku 4.5 at $1.00 per million, the gap is roughly 24x. Still a large margin, and still not the headline number. His verdict on the launch as a whole is that the speed, the cost and the novelty of the category are real, while the marketing runs about ten times ahead of the evidence — something he credits the company with half-admitting.

See Reception of Jev.

How the Latency and Calibration Claims Read Under Measurement

Two vendor figures have since been measured from outside, in Adapting Jev to Your Domain with GEPA.

Latency. TypeSafe’s stated end-to-end range is 70–500 ms. Praneeth Paikray observed a client-side median of 14.69 s (p95 15.62 s) over 505 requests, and 19.59 s (p95 24.37 s) over 1,260 evaluations at up to 24 concurrent — with 12.35 s on a serial smoke test, so concurrency is not the explanation. He is explicit that he cannot separate model inference from transport and queueing, which makes this a reading of the hosted service in September 2026 rather than of the model. It is still one to two orders of magnitude from the published range, and the company publishes no p95/p99 figures or SLA against which to reconcile it.

Calibration. The launch rests on probabilities you can threshold, trained for by RLCD. Measured at the default prompt on one classification task, Jev returned a 10-bin ECE of 0.173 against 0.052 for a TF-IDF baseline, and log loss of 1.849 against 0.335 — with confidence returning exactly 1.0 on half the test set, ten of those answers wrong. The fair qualification is that optimizing the prompt cut ECE to 0.069, so the figure indicts the default instruction at least as much as the training method. The company has still published neither a reliability curve nor an ablation isolating RLCD.

Cost. This one checks out. Input billing at 0.00898 and $0.03067 for the two experiments, exactly as advertised.

Graph View

Backlinks