OpenRouter State Of Models: Jev, Open-Weights, and Tokenomics

YouTube · Watch/Read

• Dan Disler argues there is no AI bubble, citing OpenRouter’s exponential token usage (146 trillion tokens/week, up from 4.5 trillion a year ago) as evidence of a growing “pie” where multiple players are winning.
• He highlights “tokenomics,” emphasizing that while token usage is climbing, engineers must ensure more tokens equate to more useful work, not just burning compute.
• New model classes like JEV System One (a zero-shot classifier) are gaining traction for efficiency, and “free tokens” (e.g., SpaceBunny Alpha) indicate demand for cheaper models, though “if it’s free, you’re the product.”
• A key trend is the rise of agent harness control (e.g., the Pi agent), with engineers realizing they must own their agent harness rather than just renting intelligence.
• The practical takeaway for engineers is to understand the intelligence available, scale compute to scale impact by starting with state-of-the-art models and then scaling down, and prioritize owning their agent harness.