Microsoft has folded chat, coding and long-running agents into a single Copilot app and priced it in two halves: a per-user licence for everyday work, a credit meter for everything expensive. The number that matters is not the $30 seat. It is the $43 Microsoft says an Anthropic model costs inside its bundle against the $129 it says the same model costs outside it.
McKinsey has put numbers on the thing every finance chief has noticed. A unit of intelligence keeps getting cheaper; the invoice keeps getting larger. One in five organisations now says cost is limiting what it does with AI, and the cost of the same task can vary thirtyfold between runs.
Anthropic released Claude Opus 5.5 on Tuesday and OpenAI answered within the hour with two cheaper GPT-6 models. Both labs had asked the industry to slow down on 12 September. The antitrust suit filed since can be argued either way.
Per-seat models built for human users fit poorly when software does the work, and vendors are experimenting their way toward outcome-based alternatives.
The frontier labs are converging on a similar commercial shape: consumption pricing at the bottom, committed contracts at the top, and services in between.