Jev, an AI model that returns structured decisions instead of plain text

TypeSafe AI has introduced System One, a new model category built to make fast, structured decisions that software can use directly. The first model, Jev, is available in early access. Founder Diogo Almeida says he worked at OpenAI on the methods behind ChatGPT, and the company spent two years in stealth mode.

Unlike conventional language models, Jev does not generate text. It takes unstructured input, such as the state of a program, and returns values in a format defined in advance, each one with its own probability and confidence score. It is trained with a proprietary method, Reinforcement Learning for Calibrated Decisions (RLCD), designed to match the model's stated confidence to its actual accuracy. In its charts, Jev logs a 0% error rate on structured output and tool calls, which the company says is not an empirical result but a guarantee derived from the output schema. It also claims the model cannot hallucinate, though that guarantee applies to the format of its answers and does not mean every decision is correct.

TypeSafe puts Jev's response time at 70 to 500 milliseconds, against a range of 3 to 329 seconds it attributes to current frontier models. Input tokens cost $0.042 per million, and output is free, though the company admits it cannot prove the pricing isn't subsidized.

To compare quality, it uses its own workflow evaluations, benchmarked against the average of GPT-6 Astra and Fable 5.1. In them, Jev performs on par with models such as Terra and Sonnet 5, and below Sol and Opus 5. In those same workflows, the company puts its edge at up to 193.6x faster and 444.6x cheaper, and expects real-world gains to land at the high end of what it measured. It acknowledges its own team designed the workflows and that the tests ran from laptops on the US West Coast.

Key points

• TypeSafe AI launches Jev, the first model in its System One category, in early access.
• It doesn't generate text: it returns structured values with probability and confidence.
• It's trained with a proprietary method, RLCD, to calibrate the confidence of its decisions.
• It responds in 70-500 ms, versus 3-329 seconds for current language models.
• It costs $0.042 per million input tokens; output is free.
• It logs a 0% error rate on format, a design guarantee rather than a measured result.
• TypeSafe claims it can't hallucinate, though that doesn't guarantee every decision is correct.
• In its tests, it matches Terra and Sonnet 5, and trails Sol and Opus 5.
• The company claims up to 193.6x more speed and 444.6x lower cost in those workflows.