Microsoft caps AI token budgets, switches to cheaper model

Microsoft ↗ 0.52% has told its engineers to stop "tokenmaxxing" — the practice of burning through expensive AI tokens without regard to cost or outcomes — as the company introduces division-level budgets and switches to a cheaper default model for internal use.

The Memo

The directive came in an internal email from Jay Parikh, the executive vice president who leads Microsoft's CoreAI organization, first reported by 404 Media on August 4. "Tokenmaxxing is not what we are optimizing for," Parikh wrote. "I want all of us focused on maximizing outcomes that move the needle for our customers and our business."404media

As part of the shift, Microsoft is making OpenAI's GPT-5.6 — described internally as cheaper to run — the default model for employee use. Previously, the company's internal GitHub Copilot setup had routed many requests to Anthropic's Claude models, generating costly token bills. Updated internal guidelines, effective as of July 2026, establish AI token budget targets for every division and allow individual engineers to track their own spending, which reportedly ranges from hundreds to a few thousand dollars per month.indianexpress+2

Parikh stressed that the move is not about cutting back on AI ambitions. "We are not optimizing for fewer tokens. We are optimizing for more impact per token," he wrote. The company's most recent earnings showed increases in revenue, operating income, and net income, beating Wall Street expectations.404media+1

An Industry-Wide Reckoning

Microsoft is not acting alone. ABC News reported in late July that the broader corporate enthusiasm for tokenmaxxing was already fading as companies questioned whether heavy AI usage translated into proportional productivity gains. Meta ↗ 1.02% recently imposed its own token budgets and shut down an internal leaderboard called "Claudeonomics" where employees competed to burn the most tokens. Amazon Amazon.com, Inc. ↘ 2.24%, Adobe ↗ 1.02%, Atlassian ↗ 3.41%, and Citi Citigroup Inc. ↘ 0.41% have adopted similar cost-control measures.gadgetreview+2

At a June appearance on The New York Times' "Hard Fork" podcast, CEO Satya Nadella addressed tokenmaxxing directly, telling the audience that not every problem needs the most powerful AI model.businessinsider

The Deeper Tension

The irony is hard to miss: the company that has bet tens of billions of dollars on AI infrastructure and sells Copilot subscriptions to enterprises worldwide is now rationing its own product internally. As one anonymous Microsoft employee told Fortune, it is "the ultimate admission that we, as hosts of AI infra, can't afford our own AI products". Uber ↘ 0.67% president Andrew Macdonald similarly questioned why higher token usage among his engineers had not produced a proportional increase in useful features.gadgetreview+1

The shift suggests that the AI industry's initial measure of success — sheer volume of usage — is giving way to a harder question: what, exactly, is all that compute producing?