TypeSafe AI says it's 193x faster than LLMs. Its own proof shows 75x.

TypeSafe AI says it spent two years in stealth, and it wants to be read as a rejection of the chatbot era. Its launch page opens with a line and a two-word fragment: "We took the opposite research direction," followed by "not chat."

The company, whose footer reads "Made in SF," has shipped Jev, which it calls its first public System One Model. A System One Model, per the page, is built to answer structured questions inside software rather than to hold a conversation. Three claims sit around the product: typed outputs, a confidence estimate on every decision, and "zero hallucinations." Under them is a benchmark of 193.6x faster and 444.6x cheaper than LLMs.

What TypeSafe AI actually launched

The page frames System One Models as "a new class of AI model built for decisions inside software," set against LLMs that it says "produce words for people." Jev returns typed decisions with probabilities attached, so program code can branch on them. TypeSafe says the build rests on three departures from current practice: a new architecture, a new sampler, and a new training algorithm called Reinforcement Learning for Calibrated Decisions, or RLCD. Structured output is a well-worn target by now, and the gains there tend to be measured in points, not orders of magnitude, as LFM2.5-350M's schema results show.

The page is stamped version 0.01 and copyright 2026. Alongside the announcement, it lists no accompanying paper, code release, or model card. That gap between a confident launch and an empty evidence trail is familiar: Samsung's health assistant beta launched with big promises and few specifics.

The benchmark: 193.6x faster, 444.6x cheaper

Those headline multipliers sit above a single worked example that the page labels "(proof)." In it, TypeSafe's system finishes a task in 0.114 seconds at a cost of $0.000081. The column marked "LLMs" finishes the same task in 8.566 seconds for $0.013880.

Divide the two and the numbers do not reproduce the headline. The example implies roughly 75x on speed and about 171x on cost. The page does not explain where 193.6x and 444.6x come from, and it does not define the "workflows for System One tasks" its asterisk points to. The larger figures may aggregate across a wider task set than the one shown. The page does not say that either. A speedup that shrinks once you read the attached results is a familiar pattern, and Hugging Face's 2.57x WebGPU claim came with 176 losses attached.

Zero hallucinations and the confidence claim

"Zero Hallucinations" runs as a heading. The line beneath it is narrower than the word suggests: every Jev decision carries a confidence estimate, so software can act when confidence is high and escalate for review when it is not. The page tells builders to "set the thresholds for when it acts autonomously and when it asks for review."

That routes uncertainty. It does not guarantee the model is right, and a confident wrong answer is still wrong. Agreement does not fix it, either: a panel of AI agents can agree and still be wrong when their errors correlate. The page's FAQ raises "Can Jev still get things wrong?" and "Is Jev deterministic?" among its questions, and the judgment about where to draw the line passes to whoever sets the threshold.

RLCD versus RLHF

The method claim is a fork in training. TypeSafe describes Reinforcement Learning from Human Feedback as a process that optimizes models for human preferences and produced systems "superhuman at instruction following," which it calls chat. It lists the costs: mode dropping, overconfidence, and unreliability, which it says leave LLMs dependent on a human in the loop.

Its replacement, RLCD, is defined only as "reinforcement learning for calibrated decisions." The name points straight at the failure the page identifies in RLHF, overconfidence. Whether RLCD is a genuinely new method is not something the page shows. There is no protocol, no baseline, and no comparison. The page's own FAQ asks "Is Jev just a smaller LLM?" and "How is this different from JSON mode or structured outputs?" Those are the two questions that would separate a new algorithm from a well-tuned classifier. Papers that propose a training method usually ship the caveats alongside it, the way a 33-of-36 distillation result was published with the authors' own asterisk.

Who is behind TypeSafe AI

The site links to a manifesto, a team page, and enterprise and docs sections, and it invites waitlist signups and open-role applications. The launch page names no founders, no investors, and no prior published research. Contact runs through a single address, hello@typesafe.ai, plus LinkedIn and X accounts.

Three blog posts are listed: "Introducing System One Models & Jev," "The Bitterest Lesson," and "AI: too good to be true, too bad to be useful." The second carries a summary line about compute driving progress "but what good is progress if you are not doing the right task." The third asks what comes next after RLHF-trained models that "please humans" rather than "make reliable autonomous decisions." The framing is a research agenda. The page does not show the research.

The comparison baseline problem

Jev's price is listed at $42 per billion input tokens, described as "238x lower input price than Claude Fable 5.1." One baseline, one metric, one direction: input tokens. Any headline multiplier begs the question of who it was measured against, and that holds whether the rival is named or not, as Alibaba Cloud's 9.27x memory claim against an unnamed rival shows.

Two gaps follow. A per-token price for typed decisions and a per-token price for text may not measure the same unit of work, so the 238x figure rests on two different outputs. And input price leaves out output cost, retries, and the human review the confidence mechanism is built to trigger. On this page, $42 is a price per input token, and the page stops there.

None of this makes System One Models wrong. A launch page is not a paper, and TypeSafe is early enough to be judged on what it ships next. But the loudest thing it has published is a set of multipliers, and the one block it calls "proof" returns smaller ones. Until the method behind RLCD is shown, the claims run ahead of the evidence.