Frontier AI labs may stop selling their best models. The assumption that OpenAI, Anthropic, and their peers will always meter intelligence by the token is now a liability to hold. Open-weights competition from China, profit-insensitive scaling from Google and Meta, and xAI's entry mean the frontier is no longer a two-player market. Five, six, or seven labs at roughly equal capability levels kills the high-margin API business model.
A crowded frontier commoditizes model access but not product execution. The labs with narrow leads, which describes OpenAI and Anthropic today, face a specific strategic fork: keep licensing your best models to competitors who build on top of them, or pull the ladder up and use those models to build products your competitors cannot replicate with last-generation tech. Anthropic has already signaled the direction with Claude Code and Claude Design. OpenAI is pushing the same way with Codex folded into a ChatGPT superapp.
The piece worth reading in full is not the conclusion but the logic of the calculation itself: when to sell the model versus when to sell the product built on it, and what the narrowness of your lead has to do with that choice. The real threat to enterprise software was never vibe-coding a Salesforce replacement. It was always the lab deciding its model is too valuable to hand to you at all.
[READ ORIGINAL →]