TypeSafe AI opened early access to Jev, a model the company says belongs to a new category it calls System One Models: systems built to output typed, probabilistic decisions instead of generated text. Founder Diogo Almeida, who previously worked on the instruction-following research at OpenAI that fed into ChatGPT, announced the release on the company’s blog.
The pitch is architectural, not just incremental. Where a conventional large language model samples tokens one at a time and hands back a string that downstream software still has to parse and validate, Jev generates all of its outputs in parallel and returns values matched to a schema defined in advance. TypeSafe trained it with a method it calls Reinforcement Learning for Calibrated Decisions, aimed at producing confidence scores that track actual accuracy rather than the overconfident, inconsistent estimates the company says plague chat-tuned models.
TypeSafe’s own comparison puts Jev’s end-to-end response time at 70 to 500 milliseconds, against 3 to 329 seconds it attributes to frontier models on the same class of task, a gap the company frames as a 40x to 200x speedup for what it calls “System One shaped queries.” Pricing follows the same logic: $0.042 per million input tokens, with output free because Jev does not generate text to meter. On its published workflow evaluations, which compare Jev’s structured decisions against a reference built from averaging GPT-6 Astra and Fable 5.1, TypeSafe reports gains as high as 193.6x on speed and 444.6x on cost, figures it says sit at the high end of what real-world deployments should expect.
The company’s most pointed claim is that Jev “can’t hallucinate.” That claim rests entirely on the model’s output mechanism, not on any independent evaluation of factual accuracy: because Jev’s possible answers are constrained to a predefined schema, it cannot return a value of the wrong type, so TypeSafe says it can mark its own type-error rate at zero by construction. That is a claim about structural guarantees on output format, not a claim that the model’s underlying judgments are correct. TypeSafe itself draws this distinction in its own writeup, noting the zero figure “is not empirical” but follows from schema matching being enforced rather than measured.
TypeSafe also flags several sources of bias in its own numbers. The company says its speed benchmarks were run from laptops on the West Coast, that it cannot rule out its pricing being subsidized, that its workflow evaluations were built by its own model-capabilities team, and that using OpenAI and Anthropic models as the reference answer set likely undercounts how DeepSeek’s models would compare. Those are TypeSafe’s disclosures, not third-party audits, and no independent lab has yet published a comparison of Jev against established models.
The framing matters because “hallucination” has become shorthand for a much broader reliability problem in production AI systems, one that spans not just malformed outputs but confidently wrong reasoning. A model that cannot emit a badly typed value can still emit a well-typed value that is substantively incorrect, and TypeSafe’s own methodology, benchmarking against other AI systems rather than ground truth, cannot rule that out. The company is positioning Jev for classification, routing, scoring, and similar decision points embedded inside existing software rather than as a chat replacement, a narrower and more testable claim than “cannot hallucinate” implies on its own.
For teams building agentic pipelines where a single malformed tool call can break a dependency chain, the pitch is worth a pilot: swap one brittle if-statement or classifier for Jev’s API and measure latency and cost against the current model in production, rather than taking TypeSafe’s workflow benchmarks as the baseline. Given the price TypeSafe is charging and the absence of independent verification so far, the decision to standardize on it should wait for evaluations run by someone other than TypeSafe itself.
Based on TypeSafe AI’s official announcement on its company blog, published September 16, 2026.
