TypeSafe Raises $40 Million for Jev, a Model That Won't Talk

TypeSafe AI raised a $40 million seed round led by DCVC, the firm said, on the same day the San Francisco startup came out of stealth and put its first model, Jev, into early access. Forbes reported the round values TypeSafe at $200 million, citing a person familiar with the deal; neither TypeSafe nor DCVC has confirmed that figure, which puts a specific number on a company that was founded in 2024 and stayed quiet until Sept. 15.

Chief executive Diogo Almeida co-founded TypeSafe with Erik Gafni and Sasha Sheng. Almeida spent years at OpenAI, where he worked on reinforcement learning from human feedback and the research that became InstructGPT, ChatGPT and GPT-4, according to SiliconANGLE. "We've been optimizing for humans, and we're superhuman at pleasing humans," Almeida told Forbes, arguing that is the wrong target for AI embedded inside software rather than a chat window.

A model built to skip the sentence

Jev is what TypeSafe calls a System One model, a reference to Daniel Kahneman's distinction between fast intuition and slow deliberate reasoning. Instead of generating text token by token, Jev takes in unstructured state and returns typed, structured decisions with calibrated confidence scores, all sampled in parallel rather than one token at a time. TypeSafe trained it with a method it calls Reinforcement Learning for Calibrated Decisions and says the architecture makes a malformed output mathematically impossible. End-to-end response time runs 70 to 500 milliseconds, the company said, and input tokens are priced at $0.042 per million with output free.

A price comparison nobody outside TypeSafe has checked

TypeSafe's own figures, cited by SiliconANGLE, put Jev at $0.39 per 1,000 workflows against $3.31 for OpenAI's GPT-5.6 Luna and $19.49 for Anthropic's Claude Haiku 4.5, a vendor-supplied comparison The Terminal could not independently verify.

1. Jev0.39 $
2. GPT-5.6 Luna3.31 $
3. Claude Haiku 4.519.49 $

Source: TypeSafe AI figures cited by SiliconANGLE, accessed 2026-09-22

A separate TypeSafe benchmark claims Jev finished one workflow in 0.114 seconds versus 8.566 seconds for a comparison LLM, a gap the company describes as 193.6 times faster and 444.6 times cheaper. Remio's review of the claim points out that TypeSafe designed and published the comparison itself and that no independent lab has reproduced the result. The review also flags what TypeSafe has not released alongside the number: the task definitions, accuracy rates against a baseline, or reproducible evaluation code. Speed without a published accuracy figure, the review argues, only tells a buyer how fast Jev fails when it fails.

What TypeSafe has not said

Beyond the unconfirmed valuation and the unreplicated benchmark, TypeSafe has not disclosed how many customers are using Jev, since it remains in early access, or named any investor beyond DCVC. The technical argument over whether the underlying trick is defensible is already playing out separately: Arcturus Labs has argued OpenAI could replicate Jev's core technique by formally packaging a capability large language models already have. Almeida's counter, that the moat is TypeSafe's synthetic training data rather than its architecture, is the claim the $200 million figure is ultimately a bet on. Whether that bet gets marked up the way Mirendil's seed did depends on customers TypeSafe has not yet named.