A new AI startup founded by a former OpenAI researcher has launched what it calls an entirely new class of model — one that cannot write you an essay, and says that is exactly the point.
TypeSafe AI announced on Tuesday the release of its first "System One Model," named Jev, available immediately in early access. The company was founded by Diogo Almeida, who says the research behind ChatGPT's instruction-following grew out of work he contributed to at OpenAI. After two years in stealth, he is arguing that the industry's fixation on chat has obscured where automation actually fails. For more context on this story, see our ongoing breaking AI news.
"Models have been superhuman at chat for years, so where is all the automation?" Almeida wrote in the launch post. "That has been my driving question for the last four years."
What a 'System One' Model Actually Is
The name borrows from psychology's distinction between fast, instinctive thinking and slow, deliberate reasoning. Where large language models generate text token by token — flexible, general, and prone to occasional confident nonsense — TypeSafe's Jev is built to do something narrower: take unstructured state as input and return typed, structured decisions with calibrated probabilities attached.
The company describes Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." Because the possible outputs are defined in advance and the model never generates free-form strings, TypeSafe claims it cannot produce type errors — a property the company says is mathematically guaranteed rather than statistically likely — and cannot hallucinate in the way text-generating models can.
The architecture departs from mainstream practice in three ways, according to the launch post: a new model architecture, a parallel sampler that produces all outputs in a single query rather than sequentially, and a training method the company calls Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for epistemically honest probabilities instead of human preference or programmatically verifiable outputs.
The Claims: 100x Faster, a Fraction of the Cost
The numbers TypeSafe publishes are aggressive. The company says Jev delivers intelligence comparable to existing frontier LLMs on what it calls "System One shaped" tasks — classification, routing, scoring, extraction and branching decisions — while responding in 70 to 500 milliseconds, against end-to-end response times of 3 to 329 seconds for frontier models. That works out to 40x to 200x faster, the company says.
Pricing is similarly unconventional: $0.042 per million input tokens, with output tokens free, since outputs are computed in parallel rather than generated token by token. The company acknowledged in its post that it cannot yet prove the pricing is sustainable at scale.
"Extraordinary claims require extraordinary evidence," the post reads, before offering what evidence it has.
The Receipts — and What Skeptics Say
TypeSafe published a side-by-side demo comparing Jev against GPT-5.6 Terra, which it describes as the most comparable frontier model at similar intelligence, and says the only disagreement in the recorded run involved a genuinely ambiguous judgment call. It also introduced a new "workflow eval" methodology that measures models against the consensus of the largest external models — OpenAI's Astra and Anthropic's Fable — inside a fixed compute graph, and claims Jev "owns the Pareto frontier for almost 2 orders of magnitude."
Notably, the company published an accompanying post titled "antibenchmaxxing" explaining why it avoids traditional benchmarks, arguing that a model designed for structured decisions inside software workflows resists meaningful global comparison.
The reaction on Hacker News, where the launch drew hundreds of points within hours, ranged from genuine excitement to pointed skepticism. Several commenters noted that the public evidence consists largely of demo videos — including one showing the model powering a Doom gameplay loop — rather than reproducible third-party evaluations. Others argued the approach is best understood as an extremely fast, frontier-quality classifier, useful for a slice of workloads rather than a replacement for LLMs.
Even sympathetic observers framed the burden of proof clearly. "Yes, they talk as skeptics but don't offer a ton of evidence, other than a couple videos of demos," one commenter wrote. "A live demo would be far more convincing."
Why It Might Matter Anyway
The pitch lands on a real pain point. A large share of production LLM calls in commercial software are not generative at all — they are binary judgments, category assignments, and routing decisions wrapped in prose. If a model can make those calls with calibrated confidence at 70 milliseconds and effectively zero output cost, the economics of AI-powered software change considerably: real-time user interfaces become practical, and pipeline steps that were too expensive to automate with LLMs become cheap enough to run everywhere.
TypeSafe's suggested use cases include classifying and routing data, verifying and guarding LLM outputs, detecting jailbreaks, and map-reducing analysis over large datasets — workloads where the surrounding code can constrain the model's freedom and catch errors.
The company is candid about the open questions: its published evals were run from laptops on the US West Coast, and long-term pricing sustainability remains unproven. Whether Jev's intelligence claims survive contact with independent testing will determine whether "System One Models" becomes a category or a curiosity.
Early access is open now through the company's website.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
