Microsoft's tokenmaxxing crackdown

A sign of the times: "Tokenmaxxing is not what we are optimizing for." Also: The White House has a concept of an AI plan, Apple vs. OpenAI gets nastier, Airtable's AI escape hatch, and more.

This week, Jay Parikh, the EVP who runs Microsoft’s sprawling CoreAI org, sent his teams a memo that’s designed to end tokenmaxxing, tells engineers to review their individual AI spend, and makes OpenAI’s GPT-5.6 Sol the default model for internal use. The GitHub Copilot economics underneath the memo help explain it.