Meet Jev: It is 200x faster than LLMs at decision making (and the demos are unhinged)
Jev has been the rage the last couple days.
This blog post covers what it is and aims to explain it in a non technical friendly way.
Problem: LLMs often feel slow for a lot of app work because developers force them to "write" an answer every time.
Like asking someone to write a full paragraph just to say yes or no.
Then the app has to read that paragraph, pick out the answer, and hope the model did not invent something weird.
Jev is a fresh take that skips the writing. Built by the @typesafeai team. One of them, Diogo, co-created ChatGPT while at OpenAI.
You give Jev a situation, ask a clear question with clear options, and it comes back with a decision plus how confident it is.
Jev compared with LLMs is wild:
- 20 to 200x faster
- 40 to 400x cheaper (and output tokens are free because they are too cheap to meter. What?!)
So Jev is not a better chatbot. More like a smart yes or no decision maker your app can press a thousand times.
Launch post from the TypeSafe team: Introducing System One Models and Jev
Below are the wildest demos people are shipping, with the original posts embedded so you can watch them yourself.
Speech and finger tracking to canvas clicks
@jackcheng says it plain: "Jev is the future"
Speech and finger tracking on a tldraw style canvas. Jev picks action, target, and where with confidence. The canvas clicks.
"draw a yellow circle here." Creates a yellow ellipse at 99% in 437ms. Also blue square, red diamond, "make it bigger." Latencies about 199 to 437ms.
Jev is the future pic.twitter.com/q5qLplU3Ei
— Jack Cheng (@jackcheng) September 17, 2026
Skip YouTube sponsors in real time
@tdinh_me built a Chrome extension that listens to YouTube audio, detects the sponsor bit, and skips it live.
About $0.005 per video. BYOK. Open source.
Just trying out Jev, I made a Chrome extension that:
— Tony Dinh (@tdinh_me) September 18, 2026
- Listens to your YouTube audio (optional)
- Detects if it gets to a sponsor segment
- Skips it ➡️➡️➡️
- All in real-time while costing ~$0.005 per video
Prototype project, BYOK, open-source:https://t.co/g4sPXR5kML pic.twitter.com/nH6c1LhU3C
Support ticket triage (explainer) + AI Gateway
Sean Roberts, VP of Netlify where I work, dropped an explainer on support ticket triage. This is not live support wiring.
Ticket in. Owning team, frustration score, urgency probability. Confidence gates auto handle or human.
And Jev is on Netlify AI Gateway today, callable from Functions.
Here is the full walkthrough if you want to watch it work.
— Netlify (@Netlify) September 17, 2026
A real support ticket goes in, and Jev comes back with the team that owns it, a frustration score, and a probability that it reads as urgent. The confidence value on each answer decides what gets handled automatically… pic.twitter.com/KIeiEce5Xc
Browser Use books flights
Browser Use and Jev (the clip that kicked this off for a lot of people):
Zurich to London on live Google Flights. About 7s. $0.0039. Video at 1x.
Breaking: Browser Use + Jev = Ultrafast ⚡
— Gregor Zunic (@gregpr07) September 17, 2026
Findings flights took 7s and cost only $0.0039 🤯
> new action space every step
> DOM state space
> small LLM fallback to type
(this video is at 1x speed btw)
Built a tiny open source browser agent. try it below ↓ pic.twitter.com/AplCBRYC5o
Inbox gates that make frontier models look expensive
@nutlope: 100 emails classified in 1.42s. Uncertain ones escalate to a big model. 96/100. Full pipe about $0.07.
@VaibhavSisinty: 500 emails for about 3.5¢.
Same shape showing up everywhere
- Wiki race (Coffee to AI)
- Coding agents judging tools, files, lint
- Multiplayer harnesses asking "should we reply?" every turn
- Vertical ops: reno line items, insurance docs, ticket urgency
The pattern
LLMs write. Jev decides.
State and typed questions in. Calibrated choices, scores, probs out. Then the browser, the inbox, the canvas, or the agent does the thing.
If you have more demos, drop them. I will keep stacking this list.