If you want to understand how fast AI is improving, look at intelligence per dollar. 1.5 years ago: o1 Pro: $150 / $600 per M tokens Today: GLM-5.3 Flash: $0.15 / $0.50 That’s a ~1000x collapse in price in under 1.5 years, while GLM-5.3 Flash is more intelligent than o1 Pro.

some good data here about model routing first cursor model router, now openrouter + jev. people keep trying to build routers at _request level_ which is not going to work the fundamental problem is that a prompt without surrounding context is simply not enough signal to judge complexity easiest example to understand this - if you ask “how does this work” in a toy repo with just one html file, any dumb model can do it. but if you send the exact same prompt in linux kernel repo, hell breaks loose. how do you route that prompt? the second problem is prompt caching. it means we can’t frequently change the model during a session in an economic way. one or two missteps here and there, and you find yourself spending more cost than simply using the best model all along the right way to build a router is to do it at task level, not request level. this cannot be done by a model gateway which has to assume everything is a single continuous session. it needs to be a harness layer capability which can make use of a multi-agent architecture a smart enough model first needs to do some digging and understand the complexity of the ask, then delegate substantial sub tasks to models with appropriate intelligence - even this is very hard to get right

Meanwhile at @linear

It’s been a week since I moved one of eval datasets from Gemini 3.1 to Jev by @typesafeai for LLM judging. The results so far: - No meaningful change in scoring accuracy - ~200× cheaper (~$0.01 → ~$0.00005 per judge) - ~50× faster (~10s → ~0.2s median) - ~50% fewer input tokens, 87% fewer output tokens But not everything’s perfect, more details in the thread

I spent some time investigating @typesafeai's Jev and striking to see them use a similar abstraction as autorubric -- released 8 months ago and to be presented at @COLM_conf next month. Our criterion types map 1:1 to Jev's Noul/Score/Choice: arxiv.org/abs/2603.00077

Introducing typesafe/jev-router: a cache-aware model router powered by Jev and @typesafeai The Jev Router picks the best model and reasoning effort for each request, balancing quality, speed, and cost. Here's how it works 👇🏻

Jev Router keeps a model that works for the rest of the session. It can raise or lower effort without switching models. It switches only when the expected gain is larger than the cost, including the cached chat it would lose.

I've been using Jev for all kinds of things. This morning I had a realization I kind of like: Use it to make non-black-box embeddings. Instead of an embedding model spitting out 1,536 numbers that mean nothing, you ask Jev questions about each document. The answers become the vector. An email in Cora: "I got charged twice this month, pls fix asap" [is_customer, urgent, about_billing, needs_reply] [1.0, 0.9, 1.0, 1.0] A newsletter: [0.0, 0.0, 0.0, 0.1] A friend asking about lunch: [0.0, 0.1, 0.0, 0.7] Then it's just old-school cosine similarity search. Search "billing issues from customers" as [1, 0.5, 1, 0.5] and the double charge comes out on top. Same idea for our articles at Every: [is_tutorial, about_ai, contrarian, beginner_friendly] Or support tickets: [is_bug, angry, churn_risk, enterprise] Every number has a name, so you can see why something matched. Need a new dimension? Add a question. Want urgent stuff first? Change the query vector. Trying this in @CoraComputer now to make search fast.

Introducing typesafe/jev-router: a cache-aware model router powered by Jev and @typesafeai The Jev Router picks the best model and reasoning effort for each request, balancing quality, speed, and cost. Here's how it works 👇🏻

DSPy 3.4.0 was just released! This release includes native support for Jev and System one models inside of DSPy! Use it with compatible signatures. This release also includes a brand new optimizer, ReAnchor, specifically for calibrating outputs with confidence.

everything you need to start building with jev, in one article. code, architecture, diagrams... everything you need to follow the build and make it your own.

me explaining why we can't run a classifier on every event