TypeSafe AI — typesafe.ai

TypeSafe's own benchmark comparison: 193.6x faster and 444.6x cheaper based on its workflows for System One tasks, with Jev listed at $42 per billion input tokens. The caveat says they are vendor runs rather than independent testing.

Yes. The FAQ addresses whether Jev is just a smaller LLM, how it can be so fast and inexpensive, and where it struggles; the docs' jaggedness page is the maintained version of that failure list.

The site's argument is that RLHF optimizes for human preferences, which produced models superb at instruction following but also creates inherent issues: mode dropping, overconfidence and a lack of reliability, which together mean LLMs require humans in the loop. System One Models instead pair a new architecture and a new sampler with Reinforcement Learning for Calibrated Decisions, built to be used natively by machines rather than supervised by people.

Typed outputs that software can act on, with calibrated confidence attached to every decision. The site's framing is more like code: reliable, fast, self-consistent and type-safe, against LLMs that produce words for people. You set the thresholds for when it acts autonomously and when it asks for review, then combine the decisions in code into larger workflows with control over how the intelligence is used.

A side-by-side run of the same System One task, labelled 193.6x faster and 444.6x cheaper and based on workflows for System One tasks with a link to the proof. TypeSafe records cost $0.000081 completed in 0.114 s against LLMs at cost $0.013880 completed in 8.566 s. Jev is separately listed at $42 per billion input tokens, described as a 238x lower input price than Claude Fable 5.1.

It links out to the launch post, the documentation, console sign-in, workflow evals, the blog and open roles on Ashby, alongside the product claims. Treat the page as vendor material rather than documentation: the speed, price and accuracy multipliers come from TypeSafe's own benchmark comparison, so the docs and the console are the places to check contract details like question limits.

Its own diagram runs from low-intelligence pre-trained models through RLHF chat models to RLVR reasoning models, then marks RLCD, reinforcement learning for calibrated decisions, as a separate category beside them. The contrast it draws is that LLMs produce words for people while Jev produces typed decisions that are more like code: reliable, fast, self-consistent and type-safe. Model availability, API access and pricing are documented separately on the docs site and console.