This is an opinion piece. The facts under it are sourced; the conclusion is the writer's.
The competition that matters has moved from the frontier model to the ten-cent one. Anyone still choosing a model by the leaderboard is optimising the wrong number.
The top of the market converged
Three labs, five points of index score, error bars wider than the gap. When the best models are within noise of each other, the differences that decide a product are price, latency, and whether the thing stays up.
That is a normal market. It is what happened to databases and to cloud hosting, and it took about as long.
The bottom is where the decisions are
Claude Haiku 5.5 at $0.10, GPT-6 Luna at $0.10, DeepSeek's flash tier at $0.01. A hundredfold below the flagship, and good enough for classification, extraction, routing and summarising.
An agent loop that calls a model ten times per task does not get better because each call is smarter. It gets better because nine of those calls are cheap enough to retry.
Margins say the same thing
Enterprise AI gross margins of 45 per cent rising to 53 per cent are not the margins of companies selling intelligence. They are the margins of companies selling capacity, and capacity is a commodity.
The vendors who will keep their margin are the ones whose customers cannot easily move: the ones holding the workflow, the data, or the hardware.
What I would do about it
Put the cheap model in the default path and the frontier model behind a check. Measure both. If the cheap one is within a few points on your own fifty inputs, the frontier call is a habit, not a requirement.
The frontier will keep improving. It just stopped being where the money is.