TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, announced a new class of models on September 15, 2026 — System One Models — along with its first member, Jev. The model does not generate text: it returns typed decisions with a confidence score in 70–500 ms. The post topped Hacker News with 1,715 points.
Key takeaways
- Three primitive types: Choice, Score and Noul — no free-form text
- Response time of 70–500 ms, claimed 40–200× faster than frontier models
- Pricing of $0.042 per million input tokens, with output tokens free
- Trained with RLCD instead of RLHF or RLVR
- Early access from September 15, 2026, via a waitlist
Three types instead of a string
Every model in the System One Models family takes a state and returns an answer in one of three types. Every answer carries a confidence score, so the calling code decides for itself when to act and when to escalate to a human.
| Type | Returns | Example |
|---|---|---|
| Choice | A pick from discrete options | billing · technical · account |
| Score | A calibrated numeric value | 0–2 range with decimals |
| Noul | A binary decision as a probability | 0.95 |
The documented call comes down to three elements:
RLCD: calibration instead of preference
Instead of RLHF or RLVR, the company uses a method called Reinforcement Learning for Calibrated Decisions. The optimisation target differs from chat models: probabilities are meant to match actual outcomes rather than human preferences about how an answer sounds. The mechanism itself has not been disclosed.
Symbol meaning
- …
- the confidence the model returns with its answer
- …
- the actual share of correct answers at that stated confidence
The practical meaning is simple: if the model states 0.9 confidence, it should be wrong roughly once in ten cases. The second difference is sampling — Jev produces its output in parallel within a single query instead of sequentially token by token, which is where the claimed speed advantage comes from.
Where this makes sense
TypeSafe AI targets the layer where large models are currently overused: workflow routing, conditional logic, large-scale data processing and verifying another LLM's output. The workflow evaluation figures are aggressive — 193.6× faster and 444.6× cheaper than comparable models — but the company itself flags them as the upper end of the range. Choices with a cardinality of up to 255 items are supported.
Why it matters
Classification and branching are currently the most expensive work done with the most expensive tool. If Jev's claims hold, that layer stops being a cost and latency bottleneck, and large models go back to where they have an edge — tasks that actually need language. The open question is whether typed output is enough for cases where the boundary between categories is genuinely blurred.
What's next?
- Early access opened September 15, 2026 — entry is via a waitlist, with no stated date for general availability
- TypeSafe has published no independently verifiable benchmarks and named no customers
- The RLCD mechanism remains undisclosed beyond its name and optimisation target
Sources
- TypeSafe AI — Introducing System One Models and Jev
- TypeSafe AI Docs — System One
- Hacker News — Introducing System One Models and Jev
