TypeSafe's Jev Ditches Text Generation for Typed Decisions, Claims 100x Speedups

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida — who worked on the instruction-following methods behind ChatGPT — has launched what it calls “System One Models,” a model class built to emit fast structured decisions rather than conversation. Its first release, Jev, is in early access. Instead of generating strings token-by-token, Jev outputs typed probabilistic decisions in parallel: unstructured state goes in, a schema-conformant set of probabilities comes out. Because it never produces free-form text and schema matching is enforced by construction, TypeSafe argues the model is mathematically incapable of type errors and cannot hallucinate — a property it frames as table stakes for automation, where a bad tool call buried in a dependency chain is a dealbreaker.

The company claims Jev matches frontier LLMs on these structured “System One” tasks while running roughly two orders of magnitude faster and cheaper, headlined by figures of 193.6x faster and 444.6x cheaper. Those numbers come from a custom “workflow” evaluation that scores each model against the averaged outputs of two large reference models (GPT-6 Astra and Fable 5.1) on decomposed, probability-driven decision graphs. The underlying stack pairs a new architecture and parallel sampler with a training method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD).

To its credit, TypeSafe front-loads the caveats: benchmarks were run from the team’s own laptops, the eval workflows were built in-house by its capabilities team, the reference answers bias toward OpenAI and Anthropic models, and the zero-hallucination claim is asserted by design rather than measured empirically. The bet is that many production automation tasks don’t need general-purpose text at all — just reliable, low-latency, type-safe decisions — a direct pitch against using chat LLMs as the backbone of software workflows. Demos include an AI-driven Doom bot (about $7/hour at 10 queries per second) and Wikipedia link-racing.