Article · AI translation
In just 24 days, the amount raised grew nearly 22-fold, and the valuation rose 37.5-fold.
The storm Jev stirred up has turned into real money!
Today, Jev's parent company, TypeSafe AI, announced the completion of an $870 million (about 6.308 billion yuan) Series A round, bringing its post-money valuation to $7.5 billion (about 54.375 billion yuan).
And what set all this off was just an AI model that “can't chat, only make decisions”: Jev.
The way it went viral is rather abstract, too: it doesn't chat or write code; given yes or no, it chooses “or.”
Put simply, it cuts out the generation function of traditional LLMs and focuses on one thing: making judgments.
And... because the new architecture is so cheap that it is hard to meter, output tokens are permanently free, while input costs just $0.042 (about 0.3 yuan) per million tokens.
More amusingly, someone even open-sourced a “Jev chat assistant” overnight. It can help you anticipate your boss's expectations, figure out why your partner is upset, and generate several suitable replies for you to choose from with one click:
Now that's genuinely useful.
So what is TypeSafe AI, this company that has suddenly burst into Silicon Valley's view? And can the new “decision model” field it has opened up work?
Read on and you'll find out.
After 2 years out of sight, Jev bursts onto the scene and attracts major investment
TypeSafe AI was founded in 2024 and is headquartered in San Francisco.
Before Jev went viral, it was essentially in “stealth mode” in Silicon Valley's AI circles.
Forget industry buzz: even tweets by its founder and CEO, Diogo Almeida, routinely hovered between single-digit likes and 0.
It was not until September 15 this year that TypeSafe AI emerged from stealth with its first System One model, Jev, and announced a $40 million seed round at a valuation of $200 million (about 1.45 billion yuan).
Now, as Jev has gone viral across the internet, TypeSafe AI's valuation has doubled along with it.
Investors including a16z, Sequoia Capital, and DCVC piled in, pushing TypeSafe AI's valuation to $7.5 billion (about 54.375 billion yuan) in one leap.
Meanwhile, Martin Casado, a partner at top Silicon Valley venture capital firm a16z who is regarded in the industry as a leading figure in infrastructure investing, also announced that he would personally join TypeSafe AI's board and take part in the company's next phase of strategic decisions.
And the sudden rush of capital was clearly driven by more than Jev's ubiquity on social platforms: its growth looked as though it was erupting straight out of the servers.
Just 3 days after launch, Jev's daily throughput reached 1 trillion tokens
Vercel data shows that within 24 hours of Jev's integration with AI Gateway, nearly 13% of paying teams had adopted it—more than twice the rate of any previous new-model launch—making it the fastest-adopted model in AI Gateway's history.
a16z also disclosed that Jev processed 1 trillion tokens just 3 days after launch.
By the time the Series A was announced, TypeSafe AI said about one-third of Fortune 500 companies had begun using or integrating Jev, saving customers millions of dollars in production costs.
This is rather different from a chatbot suddenly going viral.
Chatbot traffic often comes from people trying something new; after a few uses, the novelty may wear off;
But many of Jev's calls happen in the background: once software integrates it, the software has it classify, score, and make choices around the clock.
TypeSafe AI co-founder and CEO Diogo Almeida also said candidly in a Latent Space interview:
Our daily processing volume has surpassed 1 trillion tokens, and it isn't a fleeting peak; nighttime traffic continues to rise, too.
This shows that many of the calls come from machine automation, rather than ordinary users trying it out.
Our daily processing volume has surpassed 1 trillion tokens, and it isn't a fleeting peak; nighttime traffic continues to rise, too.
This shows that many of the calls come from machine automation, rather than ordinary users trying it out.
People sleep, but code doesn't.
Perhaps that is what really excites investors.
TypeSafe AI kept its usual straightforward, humorous style in its funding announcement, too. Instead of offering many grand words, it simply said it would take the things developers like further:
Build more native models for machines;
Fill out the infrastructure needed to build intelligent software;
Add the features enterprise customers have been asking for for a long time.
In response, commenters couldn't resist making a wish:
Could image features perhaps be added soon?
Could image features perhaps be added soon?
Diogo replied enigmatically with a “shh” emoji;
Presumably that means they're already working on it. (Something to look forward to.)
Perhaps the most dramatic contrast about Jev lies with its creator, Diogo Almeida.
The man who taught AI to chat turns around and stops it from talking
TypeSafe AI co-founder and CEO Diogo Almeida previously worked at Google Brain and OpenAI. He was one of the principal authors of the InstructGPT paper and also worked on RLHF, ChatGPT, and GPT-4.
In other words, he was among those who taught large models to “talk properly to people.”
But after leaving OpenAI to start a company, he became increasingly convinced that the industry was spending too much effort on making AI talk.
During the 2 years since founding TypeSafe AI, Diogo kept asking:
Models are now smart enough to write articles and code, do complex reasoning, and even solve Millennium Prize math problems;
But when models enter real workflows, why do most critical steps still need human confirmation and remain impossible to automate?
Models are now smart enough to write articles and code, do complex reasoning, and even solve Millennium Prize math problems;
But when models enter real workflows, why do most critical steps still need human confirmation and remain impossible to automate?
In the end, his answer was:
Today's AI is trained first to be a “likable assistant,” rather than a component software can call reliably.
It's good at making people feel an answer is decent, but it may not be suited to running automatically millions of times a day behind the scenes in a system.
Today's AI is trained first to be a “likable assistant,” rather than a component software can call reliably.
It's good at making people feel an answer is decent, but it may not be suited to running automatically millions of times a day behind the scenes in a system.
That explains TypeSafe AI's distinctive slogan:
Build Prod, Not God (build production tools; don't get caught up in building gods).
Diogo isn't fond of the industry's obsessive competition over benchmarks, either. He once said bluntly in an interview:
I think public benchmarks are extremely easy to manipulate. Even if developers don't deliberately cheat, they may end up being led by the rankings.
We really have been misunderstood and overlooked by investors because we don't compete on benchmarks, but I still want to keep doing it this way.
I think public benchmarks are extremely easy to manipulate. Even if developers don't deliberately cheat, they may end up being led by the rankings.
We really have been misunderstood and overlooked by investors because we don't compete on benchmarks, but I still want to keep doing it this way.
So when Jev was released, Diogo didn't put up a wall of standard rankings claiming it “beats everything.” Instead, he demonstrated Jev's practical uses and said he hoped developers would put it into their own workflows to see whether it could actually do the job.
And TypeSafe AI is clearly not a team hastily assembled around a rebellious slogan.
Of the other 2 co-founders, COO Sasha Sheng comes from Meta/FAIR, where she worked on News Feed, AI products, and research;
CTO Erik Gafni is a serial entrepreneur who founded Ravel, a multimodal AI company focused on DNA sequencing, and was an early employee at the 2 biotech unicorns Invitae and Freenome.
The broader team also brings together people from well-known tech companies including OpenAI, Google Brain, and Meta/FAIR.
Together, they are betting on a path different from “keep building bigger models”: AI may have been smart enough for a long time; what is really missing is an interface that lets software use it with confidence.
Jev's success has also shown Diogo and the entire living-room AI industry that developers really are willing to pay for another form of AI.
After all that, Diogo posted on X late at night, saying roughly:
Burning through over 1 trillion tokens a day, going without sleep for 3 whole weeks—we really need to go to bed... (dark circles)
Burning through over 1 trillion tokens a day, going without sleep for 3 whole weeks—we really need to go to bed... (dark circles)
Unfortunately, his peers clearly weren't going to let Diogo get a good night's sleep.
One More Thing
Just 2 hours after Diogo happily announced the funding news and groaned that he needed to sleep;
Microsoft chairman and CEO Satya Nadella entered the fray with a new decision model, Microsoft-Decision-1.
The model is post-trained on Qwen3.5-9B and specifically handles routing, classification, priority judgments, validation, and workflow control;
Given fixed options, it assigns a probability to each answer and supports yes-or-no and multiple-choice decisions, scoring, and evaluation of AI responses and agent behavior.
It looks almost like Jev's twin—it's so similar!
And of course OpenAI didn't pass up this attractive opportunity, either. It moved even earlier.
On October 6, 21 days after Jev's release, OpenAI opened public testing of its Decisions API.
Powered by GPT-6 Luna, it can turn text and images into 3 kinds of structured answers—Predicate, Choice, and Score. OpenAI says it can be up to 10 times as fast as calling the same model through the Responses API.
Earlier still, Cloudflare launched the open-source decision models Clef and Clef-flash on October 1. They are compatible with the Jev API, and Cloudflare even put Jev directly into its comparisons.
And all of this happened in just 24 days.
And perhaps that is what really makes Jev's latest funding round valuable.
It has done more than create a hit product: it has also made the entire industry suddenly realize that AI doesn't necessarily have to be better at talking. It can also be better at keeping quiet and diligently working in the software background.
And once the field took shape, TypeSafe AI went from having few direct rivals to facing competitors on all sides. (TypeSafe AI: A little courtesy, please?)
So Diogo probably won't get to sleep for very long...
If he wakes up in the middle of the night and glances at X, the competition will probably jolt him wide awake. (doge)
References:
[1]https://typesafe.ai/blog/series-ai
[2]https://x.com/CompleteSkeptic/status/2108594987177021737
[3]https://x.com/martin_casado
[4] https://mp.weixin.qq.com/s/TYgfqz-galSfQIBgjCEvDA
[5]https://vercel.com/blog/ai-gateway-jev-model-launch
[6]https://developers.openai.com/api/reference/resources/decisions/methods/create
[7]https://developers.openai.com/api/docs/changelog
[8]https://blog.cloudflare.com/clef-decision-models/
Feel free to share your thoughts in the comments!