SAN FRANCISCO — When Vercel’s engineering team swapped in a model they hadn’t heard of three weeks earlier, their safety classifiers ran eighteen times faster. They hadn’t rewritten a line of code. The only change was the model answering the questions.
That model, called Jev, does not generate text. It cannot write an email, explain its thinking, or carry a conversation. What it does is answer structured questions about program state with calibrated probability scores, in roughly a tenth of a second and for less than a twentieth of a cent per million tokens. Yes or no. Safe or unsafe. Spam or legitimate. The answer arrives in typed JSON, ready to act on.
TypeSafe AI launched Jev on September 15, 2026, after two years in stealth. The company is led by Diogo Almeida, who left OpenAI in 2024 after helping build the system that trained ChatGPT. Almeida co-invented reinforcement learning from human feedback, the technique that turned GPT-3 from a capable text predictor into something people wanted to talk to. He spent the two years since wondering whether that technique, which made models better at conversation, was solving the right problem in the first place.
Almeida wrote in the launch announcement that the industry had built models to talk to people, but most of the time, software doesn’t need to talk. It needs to decide. Jev is his attempt to build for that gap.
Jev is what TypeSafe calls a System One model, a classification the company invented to distinguish it from large language models. Where an LLM decodes one token at a time, building a response word by word, Jev processes a block of state and a set of typed questions in a single forward pass. It does not generate tokens; it evaluates all questions in parallel and returns structured answers with confidence scores. Because nothing is synthesized from vocabulary, the company says it mathematically cannot hallucinate. There is no generation step in which a plausible-sounding but wrong word can appear.
Still, they deserve scrutiny. TypeSafe has not released a preprint describing its training methodology, which it calls RLCD (Reinforcement Learning from Calibrated Decisions), nor has it published standard benchmark comparisons against models from Anthropic, Google, or Meta. What exists publicly are speed and cost numbers from Vercel, and testimonials from early API users describing use cases like email triage, workflow routing, and confidence scoring inside agent pipelines. These are practical results, not peer-reviewed ones. The benchmark absence matters because calibration quality, not just speed, is what makes classification models trustworthy in production.
Dario Amodei, whose September essay calling for a pause on frontier model development has focused the industry’s attention on what models should and should not be allowed to do, has not commented on TypeSafe’s approach. How Anthropic, whose Claude models are among the most widely deployed for agentic work, views a competitor model designed specifically to sit inside those pipelines as a faster, cheaper decision layer is a question not yet answered.
The $40 million that TypeSafe raised to reach this point came led by DCVC, a San Francisco-based deep-tech venture firm with previous bets in computational biology and materials science. The round had not been previously disclosed. TypeSafe opened Jev to early access on launch day; the API was briefly unreachable within hours as demand outpaced capacity, an early indicator of developer interest that TechCrunch reported in detail.
For developers, the appeal is not theoretical. An AI agent pipeline today typically routes every decision through an LLM: whether to escalate a support ticket, whether an image passes a content policy, whether a search result is spam. That reasoning costs tokens, time, and money. Jev’s pitch is that most decisions inside software are already structured. The questions are known, the options are typed, and what matters is a fast, calibrated answer. Replacing the LLM in that slot with Jev produces a measurable latency and cost reduction without changing the workflow.
Whether that use case is durable is the open question. Large language models have steadily improved on speed and cost over the past two years; the latest pricing for fast inference on frontier models is already a fraction of what they cost in 2024. TypeSafe’s margin advantage may compress as the big labs optimize their inference stacks. The company’s answer is that System One models are architecturally different, not merely faster. The industry will need independent benchmarks and a few production deployments at scale to know whether that claim holds.
For now, Vercel’s engineers have a faster classifier. Eighteen times faster, with no code changes. That is the kind of result that travels.
