TypeSafe launches Jev, a non-text decision model

TypeSafe AI has come out of two years in stealth with what it calls its first "System One Model," according to a launch post by founder Diogo Almeida. In the post, Almeida says he helped build the instruction-following methods at OpenAI that became the research behind ChatGPT.

The first public model, Jev, is in early access. It does not generate strings. Developers define the possible outputs and their structure in advance, and the model returns type-safe values, each with a probability and a confidence score. TypeSafe describes it as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." The company says this design means Jev cannot make type errors or hallucinate.

TypeSafe says the stack has three parts: a new architecture, a parallel sampler that produces all outputs in a single query, and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD). RLCD optimizes for calibrated probabilities, not for human preference or verifiable rewards.

The company's claims:

• Speed: end-to-end response times of 70–500ms, which it says is 40–200x faster than frontier LLMs on "System One-shaped" queries
• Cost: $0.042 per million input tokens, with output tokens free
• Quality: intelligence similar to existing LLMs on these tasks

Target uses include classification, routing, scoring and extraction inside ordinary code, batch processing of large datasets, real-time applications, and guardrails or jailbreak detection for LLM outputs.

TypeSafe admits some limits. It runs its published latency evals from laptops on the US West Coast. It also says it cannot yet prove its pricing isn't subsidized. The text available here does not name the benchmarks or the models behind the intelligence comparison, and it does not include independent evaluations.