Open and closed models are on different exponentials

Details

Lambert's core thesis is that the AI market is bifurcating along value curves. Where intelligence gains translate directly to productivity gains — coding agents being the first demonstrated case — users will pay premium prices for the best available model rather than settle for cheaper alternatives. The Opus 4.5 and Codex 5.2 thresholds are cited as the points where coding agents became obviously worth a substantial premium, with users reporting they would pay $2,000/month for top tools knowing quality will continue improving. On the other side, for tasks where marginal intelligence improvement doesn't change outcomes, open models will dominate through cost and deployment flexibility. Lambert argues top closed labs (Anthropic, OpenAI, with Google expected to join) will protect their best models by delaying API release to avoid distillation and preserve margins. The near-term market is distorted by massive compute supply constraints and token subsidization from continued investment flows, making true economic signals hard to read for several years. The broader implication: AI infrastructure investment and model selection should be explicitly calibrated to which side of this value curve the target application sits on.