Jev AI Just Changed AI Agents FOREVER!

Julian Goldie (0:00)

Jeff AI just made AI agents 10 times faster, and in one test, actually answered 777 questions in under 0.7 seconds for about a quarter of a cent. So that's the whole reason this model matters. It doesn't write, it decides. And because it never writes a single word, it comes back in a fraction of the time normal AI takes. So let me run you through what this actually is, because it's different from every model you've used this year. Jeff comes from a company called TypeSafe AI. They launched it on September the 15th, 2026, after two years working quietly with nobody watching. Now, the founder is Diogo Almeida. He came from OpenAI, which he actually helped invent RLF and worked on Instructor GPT. So the reinforcement learning method that made ChatGPT behave like ChatGPT was linked to him. He actually raised 40 million for this, led by DT. And what he built isn't a chatbot, can't be a chatbot. Jeff literally can't write a sentence, not one. But here's how it works instead. So with Jeff AI, you can hand it three things. A situation, a question, a list of allowed answers. You can pick one answer from your list and tell you how sure it is with this number. So for example, we have a look at this app, which you can see right here. With this app, it's basically feeding in real time my camera. Now I can give it different outfits. So for example, I'm actually wearing a blue shirt right now, but we can give it a black shirt or we can give it, for example, a purple shirt. And the AI is actually deciding for us and making very, very quick decisions, which makes the AI way faster, way faster than the normal writing model, and also a lot cheaper, right? So let's say, for example, you might give it a support message from a customer, and you might ask, okay, is this person angry? Are they confused? Are they ready to buy? And those are the only three answers it's allowed to give. Then it's going to pick one. It says, maybe, 0.81 confident, done, half a second. They call this a system one model. And that name comes from Daniel Kahneman, the psychologist. He splits thinking into two types. The system two is slow and careful, working on a sump on paper, for example. System one is actually fast and automatic.

So let's say, for example, you see a red light and then you stop. You don't write an essay about the red light. So every AR model you've used until now has been system two. It thinks out loud, word by word, even when the answer is just yes. GEV is the first one built for the red light, and this is where the speed comes from. So a normal model builds its answer one piece at a time. Each piece waits for the last one. GEV doesn't build anything. It looks at your situation once and then hands back numbers. Now, TypeSafe reports responses between 70 and 500 milliseconds. One person actually measured median at 256 milliseconds end to end. Compare that to what a frontier model takes on the same kind of question. Three seconds at the fast end and much longer at the slow end. But the speed isn't even the part that changes how you'd build things. There's something else in the docs I want to get to, and it's the reason your agents would stop feeling slow. I'll come back to that in a minute. First, the numbers, because I want you to see this properly. So TypeSafe actually published her own vow across four business workflows, security alerts, checking that AI agents finish work, invoice handling, customer service, and 711 cases in total. Now, Jav actually scored 67.8% agreement.

That took 0.4 seconds per case and cost about 400ths of a cent. GPT 5.6 Tera scored 67.9% on the same set. Basically, identical, but it took 10.1 seconds per case and cost 76 times more, as you can see in this article from Forecast. Now, Opus 5 actually scored higher, 73.1%. It's a smarter model on this task, but it took 38 seconds per case and cost 440 times more than Jav. Sonnet 5 landed at 67.8%, same score as Jav, nowhere near the same speed or price. So Jav is roughly level with mid-tier, frontier models on these decisions, about five points behind the top one, and it runs 25 to 95 times faster. Now, I have to be straight with you about those numbers, because a lot of people are sharing them without the caveat. That's TypeSafe's own test. They designed the workflows, and the correct answers weren't written by people. They came from averaging what GPT-6 astro and Claude Faber 5.1 said. So the score measures how often Jav agrees with two other AR models, not how often is actually right. TypeSafe say this openly. They also say their biggest speed and cost claims sit near the top end of what you'd see in real usage. So treat the benchmark as direction, not a promise. Now, as you can see in this article from Brian Walker, Dan Shipper's team at Everi actually ran the first real independent look at this. The head of EVALS pushed 777 judgments through Jav in under 0.7 seconds for about a quarter of a cent. Across 11 different experiments, they made 1,709 judgments for less than one cent total. They sorted startup pitches, they graded customer support replies, they flagged risky actions on agents that they were about to take, and they checked writing for AI tells. They asked 100 simulated people which advert they click on, and the verdict was honest, good, but not perfect. So Jav correctly flagged the articles that leaned more on AI, but they said they'd want a deeper accuracy check before putting it into production. Then there's a UK event site called Near Here. They tested Jav on moderated listings. Jav got 96%.