{"slug": "my-early-thoughts-on-jev", "title": "My Early Thoughts on Jev", "summary": "TypeSafe's Jev, a decision-making AI system that outputs choices, scores, and yes/no answers rather than text, is priced at $42 per billion input tokens with no charge for output tokens, according to early impressions published by the writer. The system operates at roughly the intelligence level of Sol or Opus for these decisions and returns requests in hundreds of milliseconds, with an early sweet spot around 200 ms. The writer argues Jev's near-instant, near-free classification and labeling could become a new cornerstone for AI work, including model routing, cybersecurity triage, and multi-staged eval systems that previously relied on slower, costlier large language models.", "body_md": "Okay, I’ve got a video coming out, but here are my early thoughts on [Jev](https://typesafe.ai).\n\nThe first thing to know is that it’s not a large language model. It doesn’t produce text, it outputs decisions.\n\nThe decision types are (roughly) choice, score, and yes/no.\n\nIt is roughly at the intelligence level of Sol or Opus for making these decisions.\n\nThe question is: why are people so excited about this? A couple of reasons:\n\nThe most important one is that so much of what we actually do in AI is [making these types of decisions](https://danielmiessler.com/blog/weve-been-thinking-about-ai-all-wrong). Especially for enterprise-type work and AI harness-type work where we have to do things like classification and labeling constantly. This is things like deciding if this particular email is spam or if this person with an account on your platform is about to cancel.\n\nWithin your AI harness, it's things like deciding which model you should use for a particular task. In cybersecurity, there are millions of use cases for this, where we're trying to categorize types and classifications of different inputs and the chances of something being dangerous, etc.\n\nWhen you step back and look down at most AI work, there is just a massive amount that reduces down to making these decisions at scale.\n\nAnd that's the second reason people are so excited.\n\nThis system is almost instantaneous and almost free. They're charging $42 per billion input tokens, and they are not charging for output tokens at all. That means you can make thousands or tens of thousands of requests in a lot of cases, like processing entire databases of customer interactions or processing tens of thousands of emails or whatever, and you might get a bill for like $0.12. Or maybe far less.\n\nAnd the round-trip time for making requests is in the hundreds of milliseconds. With an early sweet spot of around 200 ms.\n\nAnd when you combine these two, it means this system, or a system like it, is likely to be a new cornerstone for all AI work being done anywhere.\n\nThe first step is to zoom out and take a look at all the different work that you're doing in whatever domain or whatever application, etc. [Decompose](https://danielmiessler.com/blog/universal-business-components-ubc) how much of that work actually consists of making thousands or millions of these types of judgments and classifications.\n\nThis includes your [hook system inside of an AI harness](https://danielmiessler.com/blog/personal-ai-infrastructure), where you can do things on user prompt submit, before and after tool use, model routing, as I mentioned earlier: the classification and labeling of all your session data to find things you might need to improve or fix, etc.\n\nAnd then the next step is to essentially retool all that work so that it is going through one or more of these decision phases before handing off to an LLM. And it doesn't have to be one or the other. I have some implementations that are LLM-first, passed off to Jev for decisions, and then maybe back to an LLM, and some are doing kind of vice versa. LLMs and this type of system work really well together for tons of use cases.\n\nOne of the things I've already made massive progress on, and that has me most excited, is fully updating my eval system to include this as a layer.\n\nEvals have two branches, roughly:\n\nAsserts that are deterministic and very fast\n\nJudgment, which breaks down into rubrics and tournaments\n\nRubrics are essentially center mass for this Jev system. They are basically a set of options that AI chooses for you based on the context given and how smart the system is. Like picking how happy or upset a customer is on a five-level scale.\n\nAnd then there are tournaments which have the system pick between two options, given the context and the intelligence of the model.\n\nBut until now, these types of operations [have all been done with LLMs](https://danielmiessler.com/blog/using-the-smartest-ai-to-rate-other-ai), which are very slow and very expensive.\n\nWhat I, and many others I'm sure, are building is a way to have multi-staged evals where a vast majority of the work can be done by this system, basically instantly and for free, while shoveling out to an LLM if it is specifically needed.", "url": "https://wpnews.pro/news/my-early-thoughts-on-jev", "canonical_source": "https://danielmiessler.com/blog/early-thoughts-on-jev?utm_source=rss&utm_medium=feed&utm_campaign=website", "published_at": "2026-09-19 04:38:00+00:00", "updated_at": "2026-09-19 05:53:05.816590+00:00", "lang": "en", "topics": ["ai-products", "ai-tools", "artificial-intelligence", "ai-infrastructure"], "entities": ["TypeSafe", "Jev", "Sol", "Opus"], "alternates": {"html": "https://wpnews.pro/news/my-early-thoughts-on-jev", "markdown": "https://wpnews.pro/news/my-early-thoughts-on-jev.md", "text": "https://wpnews.pro/news/my-early-thoughts-on-jev.txt", "jsonld": "https://wpnews.pro/news/my-early-thoughts-on-jev.jsonld"}}