Jev AI Model: Inside TypeSafe’s $40M System One Bet

The Jev AI model went into early access on September 15, 2026, the same day its maker walked out of two years of silence. Behind it, TypeSafe AI raised $40 million and built something that returns a typed answer with a probability attached, never a sentence. However, the numbers behind it all come from the company itself.

At its core, Jev is what TypeSafe calls a System One model. Give it context and a fixed question, and it hands back something a program can act on directly: a category, a ranking, a score, a yes or no with a probability beside it. There is no paragraph to parse.

That framing is the whole pitch. Chat models were built to satisfy people, and TypeSafe argues that software needed something else all along. As CEO Diogo Almeida put it to SiliconANGLE, “We’ve been optimizing for humans, and we’re superhuman at pleasing humans.”

Conventional language models write one token, then the next, then the next. Jev does not. Instead it scores every candidate value at once, using what TypeSafe describes as hardware-aware parallel sampling. Because of this, the answer arrives in one pass rather than a stream.

The model handles up to 255 discrete options through a two-stage scoring process. Meanwhile, a training method the company calls RLCD, short for Reinforcement Learning for Calibrated Decisions, is meant to keep the stated confidence honest. In other words, when Jev says it is 80% sure, TypeSafe wants that to be true roughly 80% of the time. Developers can then set their own threshold for when a human steps in.

Above all, the pitch lives or dies on latency. TypeSafe reports response times between 70 and 500 milliseconds, against three to 329 seconds for conversational models on comparable work. In one demo the company showed 0.114 seconds where OpenAI’s GPT-5.6 Terra took 8.566 seconds.

Jev Against a Conventional Model

Here is how the company frames the gap, measure by measure.

On workflow benchmarks TypeSafe claims speeds up to 193.6 times faster than GPT-6 Astra and Fable 5.1. SiliconANGLE cites an internal test at roughly 194 times faster and 445 times cheaper. Crucially, nobody outside the company has verified any of it.

Pricing is where the story gets concrete. For a start, input costs $0.042 per million tokens, and output tokens carry no charge at all, because the outputs are structured values rather than generated prose. Conversational models sit between $0.20 and $10 per million input tokens.

Access runs through a hosted HTTP API, with official Python and JavaScript SDKs. Still, you cannot simply sign up today. TypeSafe opened developer access on September 16 and is pulling people off a waitlist over time. As a proof of concept, the company ran a Jev bot through the 1993 game Doom at ten queries per second, which worked out to about $7 an hour.

TypeSafe AI is a San Francisco company, founded in 2024 and quiet until this month. Meanwhile, investors led by DCVC put $40 million into the seed round, and Forbes reported a valuation near $200 million.

Diogo Almeida runs it as CEO. He worked at OpenAI on reinforcement learning from human feedback and co-authored the InstructGPT paper that ChatGPT was built on, which makes his move away from chat the interesting part. Erik Gafni and Sasha Sheng co-founded the company with him. In addition, the raise lands in a year that has already been generous to model labs, so investor appetite has clearly not cooled.

The use cases are narrow on purpose. In particular, TypeSafe points at customer service routing, tool selection inside AI agents, large-scale data classification, and real-time feature extraction. Moreover, it suggests Jev as a verification layer that checks another model’s output before anything ships.

That last one is telling. A slow, expensive chat model does the creative work, and a fast, cheap decision model polices it. For anyone already wiring up agentic AI frameworks, that is the gap Jev is aiming at: the dozens of small judgement calls an agent makes between the interesting ones.

Two caveats deserve space. First, every benchmark above is internal. TypeSafe ran the tests, published the figures, and no independent lab has repeated them yet. Early-stage labs rarely publish numbers that flatter a rival.

Second, the hallucination framing is softer than it sounds. Thomas Claburn at The Register notes that calling Jev hallucination-free means little when the model never produces natural language in the first place. A typed output cannot invent a citation, yet it can still be wrong, and a confidently wrong routing decision is its own kind of problem.

If the agent layer is what pulled you in, compare the best AI coding tools and see what developers are actually shipping with today. For another look at how an AI assistant gets priced and packaged, read our report on the Instinct AI assistant.

Sources

SiliconANGLE, The Register, heise online, AI News.*Photos: BalticServers data center by BalticServers.com, CC BY-SA 3.0, cropped and resized; Pioneer Building, San Francisco by HaeB, CC BY-SA 4.0, cropped and resized.*