I was as excited about TypeSafe AI’s Jev as anyone. Last week I built Jev Pong, not because Pong matters but because I’m in the realtime business and Jev looked like something new: a general model that makes a typed decision in a few hundred milliseconds, for almost nothing, with no training. Put that in a loop with a person and you can make decisions at a rate nobody could afford before. The post on LinkedIn did better than anything I’ve put out in years.
One comment stuck with me, because I’d been asking it myself: what’s the first real use outside a game? I said I’d follow up with a post on use cases, and this is it. I went looking for the answer this week and I expected to find it.
What I tried
I did what I’d do for any customer. I took the realtime products we carry at Ably, chat, live events, agents talking to people, and tried to find Jev a place in them.
Moderation was the obvious one. Every chat product needs it, and a cheap model that can judge a message before it lands sounds perfect. Then I remembered who does moderation for a living: specialized providers that understand harm across languages and cultures, run review queues, handle appeals and enforcement. A probability from a model is the easy part. The model was never the problem.
Trading was next, and it was the worst of my ideas. Jev isn’t fast enough for anyone who trades at speed and isn’t accurate enough for anyone who trades for real. Games, which is most of what people built, are lovely demos. I built one, and a specialized model would beat Jev on latency and cost.
So I stopped guessing and looked at the data. I sorted 5,950 posts from Jev’s first week, every post in the main Jev feed, labeled against a written rubric and audited blind by a second model. Then I read what the two public gateways show and counted the rivals. The full report, with every chart, the data and the audit, is at mattheworiordan.github.io/jev-landscape. Here’s what it told me.
What the data says
Jev on the two public gateways in its first week. Source: the report.
Jev is being used, and the money is tiny. On 23 September it was a quarter of all requests on Vercel’s AI Gateway, and 2% of its tokens. On OpenRouter it served more requests that week than GPT-5.6 Luna, for an eighth of the spend. TypeSafe hasn’t published a usage number, and 96% of OpenRouter’s Jev traffic comes from apps that don’t say who they are. A huge number of very small calls, and almost nobody who can say what they’re for.
What TypeSafe announced, what people claimed, and what anyone measured. Source: the report.
The launch numbers, 40 to 400 times cheaper and 20 to 200 times faster, are against frontier models, and nobody sends a yes-or-no question to a frontier model. Against the small models you’d actually use, the published head-to-heads I could find put the median at 7x on cost and 5x on latency, with accuracy that matched or beat the small model in most of them. That’s worth having, but it isn’t a different kind of software.
The 91 builds most likely to show something new, and what would have done the job before Jev. Source: the report.
Four in five posts compared Jev with nothing at all. Of the 91 builds most likely to show something new, the ones where someone measured a decision in a live system with a person waiting, not one did something that was unavailable before. An LLM already did 64 of them. Rules did 12. A vendor API did 11.
Then came the part I hadn’t expected. Twenty rivals appeared within the week. One was trained by agents in twenty hours on a single GPU for about three thousand dollars. One was a Shopify engineer over a few days. None matches Jev’s mix of cost, speed and accuracy yet, but the moat is a lot smaller than the launch suggested.
Where people built, against how soon the decision is needed. The cells are the labeling model’s calls. The audit confirmed only the games numbers: 11% of posts need a decision under 300 ms, and 91% of those are games. Source: the report.
Where that leaves Jev
If you want the speed, you’re really talking about games, where decisions have to keep up with frames. The reaction-rate cases everyone is excited about, a voice agent deciding whether to speak, a chat message held before it lands, a human taking over mid-conversation, are 4% of what got built, and none of it is measured in production yet. Everything else is classification. Cheap classification is useful, but if you do it at scale you’ve had a classifier for years, and if you don’t, a small LLM was already cheap enough.
That’s the surprise. The sweet spot is far narrower than it looks, and far narrower than I thought a week ago.
So, my conclusion. Jev is awesome tech for tinkerers. Generalized intelligence at low cost and low latency, in an afternoon. That’s real, it did not exist before for the masses, but its application is far narrower than the hype.
My bet is that the open-weights specialists and a Haiku-shaped decision model from a frontier lab replace it shortly. Jev the product probably won’t survive that. Jev the idea will, and I’m glad it happened. Low-latency general decision models are brilliant for prototypes, PoCs and narrow use cases, and Jev pushed a lot of us to think about what a decision should cost and how fast it should be.
Think I’ve got this wrong? The data’s open. Do tell me!
Disclosure: I run Ably, a realtime infrastructure company, which is why the live tier is the one I watch.