TypeSafe AI, founded by Diogo Almeida, announced the release of Jev, its first System One Model, positioning it as a new frontier in AI optimization for automation rather than chat applications.
According to the company, existing large language models like ChatGPT have achieved superhuman capabilities in conversational tasks but have not translated into widespread automation. TypeSafe built an entirely new stack to address this gap, introducing a novel model architecture, a parallel sampler for efficiency, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD).
Unlike traditional LLMs that generate strings of text, Jev produces type-safe structured outputs with calibrated confidence scores. The company claims this eliminates hallucinations and type errors—stating it is “mathematically impossible” for the model to make type errors. Jev is optimized for tasks like classification, routing, scoring, and data extraction, where structured decision-making can integrate directly into software workflows as decision rules.
Performance metrics, according to TypeSafe, show significant advantages over existing frontier models. The company reports end-to-end response times of 70-500 milliseconds for Jev compared to 3-329 seconds for existing frontier LLMs, representing 40x-200x faster performance on equivalent intelligence tasks. Pricing is also differentiated: TypeSafe charges $0.042 per million input tokens, with output tokens offered at no additional cost, compared to $0.20-$10 per million tokens for existing models.
The critical distinction lies in sampling methodology. Traditional LLMs generate outputs sequentially, one token at a time, while Jev generates all outputs in parallel in a single query. This architectural difference enables Jev to maintain consistent confidence estimates—a capability TypeSafe argues is essential for reliable automation, since models unable to accurately identify when they may fail cannot be trusted to run autonomous workflows.
Typeface presented workflow evaluations comparing Jev against large external models including GPT-5.6 Terra on multi-step decision tasks representative of production workloads. The company also released a side-by-side demonstration comparing outputs on structured queries.
Jev is available in early access as of September 15, 2026. TypeSafe notes that the company spent two years in stealth development and faced numerous technical challenges before reaching this announcement.
Key facts
- Jev generates structured, type-safe outputs rather than text strings, eliminating hallucinations and type errors
- Response times range from 70-500 milliseconds, 40x-200x faster than existing frontier models
- Input token pricing is $0.042 per million tokens with output tokens at no additional cost
- The model uses Reinforcement Learning for Calibrated Decisions (RLCD) training methodology
- Parallel sampling generates all outputs simultaneously rather than sequentially by token