Microsoft caps engineer AI token spend, names GPT-5.6 default

There is a memo doing the rounds inside Microsoft that puts a real ceiling on something the rest of the industry has mostly been whispering about. Engineers there have been spending anywhere from hundreds to a few thousand dollars a month on AI tokens, and the company has decided that is enough. 404 Media reported that Jay Parikh, an executive vice president, told staff by email that 'tokenmaxxing is not what we are optimizing for' and that divisions will start operating against an AI token budget target from July 2026, with individual usage tracked.

The framing Parikh chose is careful. He is not telling engineers to use less AI, he is telling them to get more out of what they use. 'We are not optimizing for fewer tokens. We are optimizing for more impact per token,' he wrote, and the practical version of that policy is making OpenAI's GPT-5.6, which is cheaper than the alternatives, the default model for internal work. If an engineer wants the more expensive stuff, presumably that is going to need a reason.

Why this matters beyond Microsoft: this is the company that sells GitHub Copilot to everyone else. When its own EVP is publicly warning that internal token consumption has to be reined in, every CFO who signed a Copilot enterprise agreement is going to have questions. According to the reporting, Amazon, Adobe, Atlassian, and Citi are pulling on the same lever, so this is not an idiosyncratic Microsoft moment, it is a shift in how big enterprises are buying AI.

The honest caveat is that the memo does not put an actual dollar figure on the budget, does not say how 'impact per token' will be measured, and does not tell us which internal workflows will be forced down to GPT-5.6 versus which get to keep the premium tier. Those details matter, because a poorly calibrated cap turns into an engineering-productivity tax pretty quickly.

The winners here, if this is the direction the industry is going, are the cheaper models that get promoted to default and the FinOps tooling that quietly meters all of it. The losers are anyone whose product roadmap assumed enterprises would keep buying tokens the way they used to buy cloud credits.