Microsoft has set fixed upper limits for AI token consumption for its business units starting July 2026 and is making the cheaper model GPT-5.6 the internal standard. According to an internal memo from CoreAI chief Jay Parikh, the reason is a significant increase in costs since developer teams have increasingly been using agentic tools like GitHub Copilot. Individual engineers reportedly spent several thousand dollars per month on tokens previously.
Microsoft sets up token budgets for each business unit
Parikh, as Executive Vice President, leads the CoreAI division, which is responsible for Microsoft’s developer tools – including Copilot and GitHub – as well as a large part of the internal AI infrastructure. According to 404 Media, which first published the email, he sent it this week to engineering teams within the company. Each division will reportedly receive its own token budget as a target figure, and individual employees will also be able to track their consumption via an internal dashboard.
Parikh writes literally, “Tokenmaxxing is not what we are optimizing for” – consuming as many tokens as possible is not the goal. Instead, the workforce should focus on results that matter for customers and the business.
As another cost-saving measure, OpenAI’s GPT-5.6 becomes the default setting for internal requests because it is reportedly cheaper than previously used alternatives. Already in July, the company had begun redirecting some requests in Excel and Outlook to its own MAI models to cut costs for models from OpenAI and Anthropic. Both steps aim at the same goal: spending less money per completed task without cutting overall AI usage.
Costs soared despite cheaper token prices
Individual developers reportedly spent between several hundred and a few thousand dollars per month on tokens, according to the memo – independently unverified. The cause is primarily the shift from simple autocomplete suggestions to agentic tools that independently execute multiple work steps in a row, consuming a multiple of the tokens in the process. According to the magazine TheNextWeb, token prices have fallen by roughly 98 percent since late 2022, while many companies’ AI bills tripled over the same period – higher volume, in other words, fully eats up the savings per request.
Parikh stresses in the memo that the goal is not fewer tokens, but more impact per token. An anonymous Microsoft employee told 404 Media the move feels like an admission that the company can barely afford its own AI products internally. For workforces outside Microsoft, the case offers a sober takeaway: teams rolling out agentic coding tools should track cost per solved task rather than just the token price – otherwise similar surprises may follow, as they did for Microsoft’s own engineers. Microsoft CEO Satya Nadella had already warned of such a cost trap in AI usage back in July.
Other companies are cutting their teams’ AI spending too
Microsoft is not the first company to limit its employees’ AI usage. Tesla capped its employees’ AI spending at $200 per week back in July 2026, after internal leaderboards had previously fueled consumption. According to the same report, Uber, Meta, Amazon, and Walmart had already introduced comparable spending limits. TheNextWeb also reports that Adobe, Atlassian, and the bank Citi have cut their internal AI budgets as well.
Externally, Microsoft continues to promote Copilot without reservation as a tool every company should roll out to its entire workforce – the same message the company has used for months to market its own AI products. Internally, the opposite motto now applies: not more consumption, but more impact per token spent. According to TheNextWeb, the cost-cutting has not dented the business figures so far: Microsoft’s AI business keeps growing sharply, and the memo is aimed solely at internal cost control.
What matters now is whether the internal austerity squares with Microsoft’s external message that every company should deploy as many Copilot licenses as possible. If “impact per token” becomes the internal yardstick, it could become a model for other companies also struggling to keep their AI budgets in check. It also remains open how Microsoft intends to convince its own developer teams to use more sparingly a tool the company simultaneously markets as indispensable.
