TypeSafe AI exits stealth with $40M and a model that writes no text

TypeSafe AI left stealth on 15 September with $40 million in seed funding and Jev, a model that returns typed decisions rather than sentences. DCVC led the round, and SiliconANGLE reported that Forbes put the valuation at $200 million, citing a person familiar with the deal. The company calls Jev the first of a class it labels System One Models, because it only decides rather than writes.

TypeSafe was founded in 2024 by chief executive Diogo Almeida, who worked on reinforcement learning from human feedback, InstructGPT, ChatGPT and GPT-4 at OpenAI. Co-founders Erik Gafni and Sasha Sheng build with him in San Francisco. Their claim is that the training which made models good at pleasing people made them awkward for software to call, a tension we traced in how RLHF works and what it costs the model.

Jev takes structured program state and hands back three output types, according to the launch post. There’s Noul, a yes or no with a probability, Choice, one option from a defined set, and Score, a rating on a scale. Every answer carries a probability and a confidence figure, so you can set the threshold where a workflow acts alone and where it escalates. Input costs $0.042 per million tokens, or $42 per billion, and output tokens are free.

While Jev gives up string generation, it’s optimized for structured outputs and can’t hallucinate.

Diogo Almeida, TypeSafe AI

That last claim rests on the output format, not on better knowledge, because a model picking among predefined values can’t invent one that isn’t in the set. It’s a narrower promise than the word hallucination usually carries, which is part of why the word causes trouble.

The launch drew a crowd, and not a quiet one. Latent Space noted the post “sat comfortably atop Hacker News all day” and clocked Almeida’s announcement on X at 4.21 million views.

Model, workflow mode	Accuracy	Cost per case	Seconds per case
Jev	67.8%	$0.0004	0.4
sol	74.1%	$0.0836	23.3
opus 5	73.1%	$0.1761	37.8
terra	67.9%	$0.0304	10.1
sonnet 5	67.8%	$0.1174	78.1
luna	66.8%	$0.0033	12.9
haiku 4.5	53.6%	$0.0195	12.5

Those rows come from TypeSafe’s own workflow evals site, which averages four automation tasks. Read them closely and the pitch isn’t accuracy. Jev ties sonnet 5 at 67.8% and trails sol at 74.1%, but it holds that tie at roughly 293 times lower cost and 195 times less time, by our arithmetic on those two rows.

That gap matters, so it helps that the company publishes the reasons to doubt it too. Reference answers are the average of GPT-6 Astra and Fable 5.1, which TypeSafe says biases the comparison toward OpenAI and Anthropic models. The four workflows came from its own model capabilities team, so it concedes “some bias could exist”, and the timings were run from laptops on the West Coast where the service is hosted. SiliconANGLE noted the speed and cost figures haven’t been independently verified.

One of the biggest remaining challenges in AI.

James Hardiman, General Partner, DCVC, on what TypeSafe is addressing, via SiliconANGLE

Still, some of it you can check yourself. TypeSafe published a Python adapter on GitHub under an MIT licence that swaps its client for OpenAI or Anthropic calls, so the same harness runs against either. Whether the confidence scores hold on your data is the part only your data answers, and the cost maths shifts if they don’t.

The framing has company in research. A position paper posted to arXiv on 18 September argues for a foundation model operating system that, “like the human brain switching between fast intuition and slow deliberation”, learns when to intervene and when to let inference proceed. Jev sits behind an early access waitlist for now, so the thing to watch is whether its confidence scores stay calibrated outside the four workflows its makers chose.