Microsoft has introduced AI token budgets for its divisions and instructed engineers to use OpenAI's GPT-5.6 Sol as the default model in GitHub Copilot, according to an internal email first reported by 404 Media and confirmed by the company to CNBC. Executive Vice President Jay Parikh wrote that the company now wants to manage coding AI "with the same discipline we apply to every other critical resource," shifting focus from maximum token use to "maximizing outcomes that move the needle for our customers and our business." The move comes after Microsoft spent months encouraging engineers to adopt AI coding tools, only to find that spending can run from hundreds to several thousand dollars per engineer monthly without clear productivity gains.

Every Microsoft division has operated under an "AI token budget target" since July, internal Copilot guidelines reveal. Workers can now monitor their personal token consumption, while the company has not disclosed budget sizes or whether engineers automatically lose access when limits are reached. The guidelines warn that individual divisions may impose additional restrictions as Microsoft tracks expenditures. Sol remains the priciest model in OpenAI's GPT-5.6 lineup at $5 per million input tokens and $30 per million output tokens, compared to Terra's $2 and $12 rates following July 30 price reductions, and Luna's $0.20 and $1.20 charges. A Microsoft study of tens of thousands of engineers found that those who adopted Claude Code and GitHub Copilot CLI merged roughly 24% more pull requests than researchers estimated they otherwise would have during a four-month period, though the research didn't establish whether the extra code reduced bugs, improved security, saved time, or delivered more customer value.

Parikh emphasized in the memo that "Tokenmaxxing is not what we are optimizing for," adding that the company seeks "more impact per token" rather than fewer tokens overall. The guidelines note that Copilot's automatic model selection had sometimes directed Microsoft employees to Anthropic models, according to CNBC reporting. Microsoft already began consolidating its internal coding infrastructure in May, when it started eliminating most Claude Code licenses in its Experiences and Devices division and told engineers to migrate to GitHub Copilot CLI by June 30, The Verge reports.

The challenge intensifies as coding tools evolve beyond autocomplete functionality, the report explains. Agents can now explore repositories, construct implementation plans, run commands, execute tests, and revise their own output, with each step adding context and output tokens to costs. An agent that repeatedly scans a large codebase or pursues an unsuccessful strategy can burn through substantial resources before a developer steps in. GitHub activated usage-based billing and user-level budget controls on June 1, citing pressure from longer agent runs, multistep tasks, and more capable models on included usage allowances. Other major companies face similar pressures: Uber reportedly exhausted its entire 2026 AI coding budget in the first four months of the year and has since shifted to cheaper default models and better spending visibility, while an internal Amazon Claude Sonnet project intended to match author records with product listings cost $1.8 million—exceeding its planned budget by 860%—and never shipped. The industry appears to be moving past simply using more AI as a measure of progress, with companies now required to demonstrate that their spending produces measurable results. Organizations that set budgets without understanding which tasks justify premium models risk either stifling innovation through overly tight controls or perpetuating waste by treating all AI work as equally valuable.