Andreessen Horowitz led an $870M Series A in TypeSafe AI on 9 October at a $7.5B valuation, 24 days after the San Francisco company announced a $40M first institutional round. Sequoia Capital and the existing investor DCVC joined, and Martin Casado, a general partner at Andreessen Horowitz, took a board seat.
TypeSafe sells a model called Jev that answers set questions about a piece of text and attaches a probability to each answer, instead of writing a reply in words. Diogo Almeida, who worked at OpenAI from 2020 on the training methods behind InstructGPT and ChatGPT, founded the company in 2024 with the former Meta research engineer Sasha Sheng and the engineer Erik Gafni. Almeida is chief executive.
The seed round in September valued TypeSafe at around $200M, according to Forbes. The Information reported days later that the company was in talks to raise more than $1B at above $10B, so the round that closed came in below the figure under discussion.
Typed Answers, Not Sentences
A developer sends Jev a block of text or structured data along with the questions to be answered about it. The model returns one of three answer shapes: a choice from up to 255 options, a score on an ordered scale, or a yes-or-no probability. Every answer arrives with a confidence number attached.
A developer sets a threshold against that confidence number, so the software acts on its own above the line and hands the case to a person below it. That is the same pattern that governs where software may act without a human watching. The jobs listed in the company's documentation are routing support messages, screening invoices, triaging security alerts and checking the output of other AI systems.
Jev evaluates every question in a single parallel pass rather than generating one word at a time, which the company gives as the source of its speed. InfoQ reported that the model shares the underlying architecture of a chatbot without being one, and that TypeSafe trained it on computer-generated examples rather than real ones, using a method the company calls reinforcement learning for calibrated decisions.
Pricing starts at $0.042 per million tokens of input, the chunks of text a model reads, with output charged at nothing. Jev holds 32,000 tokens at a time, answers in 70 to 500 milliseconds by the company's measure, and takes text only, so it cannot read images.
TypeSafe named the model after William Stanley Jevons, the Victorian economist who observed that making a resource cheaper tends to increase how much of it gets used. It markets Jev as a System One model, borrowing Daniel Kahneman's term for fast, intuitive judgement, and Andreessen Horowitz leaned on the same idea in its announcement: "make intelligence cheap enough to call anywhere, and it gets called everywhere."
James Hardiman, a general partner at DCVC, framed the September investment around reliability rather than capability. TypeSafe is "approaching one of the biggest remaining challenges in AI", he said, "turning increasingly capable models into technology that developers can reliably build into products at scale".
OpenAI Shipped A Rival In Three Weeks
OpenAI opened a Decisions service for public testing on 6 October, three weeks after Jev reached early access on 15 September. It runs on GPT-6 Luna, offers the same kinds of answer shape, charges $0.10 per million tokens of input with output free, and accepts images.
Nikunj Handa of OpenAI credited Jev with inspiring the product on the Latent Space podcast, telling listeners it had not been on the roadmap before Jev appeared, according to Fortune. TypeSafe's answer to it is a lower price and a head start of a few weeks.
Google and Anthropic reach the same work through their general models. A chief technology officer quoted by InfoQ found Gemini slightly more accurate on email classification but 10 to 20 times more expensive, and developers described pairing Claude with Jev to handle the cases Jev flags as uncertain.
The Claimed Speedups Shrink
TypeSafe says Jev runs 193.6 times faster than frontier models, meaning the largest general-purpose systems, and puts the cost saving at 444.6 times. It concedes in the same documents that those numbers sit "on the higher end of real world gains". Its tests ran from laptops on the American west coast, the reference answers came from OpenAI and Anthropic models, and its own capabilities team built the workflows being measured.
OpenChamber analysed 12,759 posts from the launch and found a median user-reported speedup of 7 times, well short of the headline figure. Median cost saving came out at 30 times and median response time at 76 milliseconds.
The company documents its own weak spots, listing unreliable counting, arithmetic and date comparison, along with accuracy loss on large or noisy inputs, and advising developers to keep maths in code. Commenters on the developer forum Hacker News noted that a model which cannot produce an invalid answer shape can still produce a wrong answer in a valid shape.
Armin Ronacher, chief technology officer of Earendil, put the design question plainly to TechCrunch. "At the end of the day, it delegates the hallucination problem a little bit to the user," he said, adding that somebody still has to decide whether a 50% probability is worth acting on.
Adoption Figures That Disagree
TypeSafe says a third of the Fortune 500 now use Jev. Andreessen Horowitz, announcing the round on the same day, put the figure at 25%, and neither has reconciled the two.
The one adoption figure attested by a third party came from Vercel, which routes developer traffic to AI models and said about 13% of its paying teams had called Jev within 24 hours of launch. Vercel described that as more than twice the share any previous model launch had reached on its platform.
Everything else rests on the company. It claims a trillion tokens processed in three days and more than a million users within weeks, and it has published no revenue figure. Some observers quoted by TechCrunch suspect that a freely available language model sits underneath Jev, a suspicion TypeSafe has not addressed either way.
A Markup Done In 24 Days
Almeida describes the work as fixing a mismatch rather than building a bigger model. "We've been optimizing for humans, and we're superhuman at pleasing humans," he told Forbes, and he told TechCrunch that "computers speak a different language". He has also said that making the model reliable took two years against the week he had expected.
Gartner reported in May that worldwide spending on artificial intelligence would reach $2.59T during 2026, growing 47% year on year, with the slice spent on models themselves at $32.6B. TypeSafe's valuation moved from about $200M to $7.5B inside that market in 24 days, with no revenue disclosed, one third-party adoption figure on the record, and OpenAI now selling a competing product at the same layer.
