Microsoft is tightening the screws on internal AI spending. The trigger: a salary-sharing spreadsheet employees use picked up a new column for monthly AI costs.
Business Insider got the spreadsheet. It contained self-reported data from about 350 U.S. employees. The biggest number on it was $28,000 over 28 days, or about RMB 190,000.
Android Headlines reported that Microsoft has started putting token use on a tighter leash through department budgets, dashboards that show each person’s token spending, and a change to the default internal model: OpenAI’s GPT-5.6 Sol.
Put simply, the company has started auditing AI spending inside its own walls.
AI spending moved next to pay data
The weird part of the spreadsheet was not just the raw numbers. It was the placement. Employees were listing AI spending right beside salary and bonus data. So monthly token use became something everyone could see and compare.
And these were not back-of-the-envelope estimates. The report said Microsoft gave employees an internal tool to check what they had spent on AI over the prior 28 days. Across the voluntary submissions, the median was $300 a month, roughly RMB 2,000.
By itself, that number was not all that shocking. The article pointed out that the fully loaded monthly cost of a Silicon Valley software engineer often starts around $20,000 to $30,000, so a $300 AI tab looks tiny next to that.
The real gaps showed up across teams. Business Insider’s department breakdown put the median for CoreAI, Microsoft’s main AI engineering group, at $975, about RMB 6,577, or more than triple the companywide median. Microsoft AI was about $490. Experiences and Devices was about $250. Azure was about $241. Several departments had individual peaks above $10,000. In CoreAI, one employee hit $16,000. The person with the $28,000 record over 28 days came from Customer and Partner Solutions.
Still, the sample was limited. Those 350 employees made up only 0.16% of Microsoft’s 223,000 global employees, and the entries were voluntary, not part of any companywide survey.
$28,000 in 28 days comes out to roughly $1,000 a day. The article tied that to an earlier estimate from GeekPark’s piece titled "Microsoft Hits Pause on Vibe Coding," which said the daily cost of a software engineer making $300,000 a year was a little over $800. On that math, the token bill for this one employee had climbed past the daily cost of hiring one more engineer.
“Tokenmaxxing” has become a recognizable behavior
In Silicon Valley, the article said, this behavior already has a name: tokenmaxxing.
That can mean intentionally sending bloated prompts, cramming context windows full, or running automated queries at scale. The point is not always solving a problem. Sometimes it is just making the usage number go up on internal dashboards.
A label like that does not appear out of nowhere. It shows up when enough people recognize the pattern. Over the past year, employees across tech have gotten a similar message: using AI proves you are ahead, and people who use it well will replace people who do not. Once management keeps hammering AI, and usage turns into a visible score, pushing the score higher becomes an easy bit of internal signaling. Quality of output? Arguable. Usage? Easy to count.
The article cited a few recent cases. Uber’s CTO told The Information that the company used internal rankings to push more AI usage, and the annual AI coding budget was gone in four months.
Meta’s version was bigger. Much bigger. In April, an employee built a dashboard called Claudeonomics that ranked 85,000 coworkers by token use. Over 30 days, the company burned through 60 trillion tokens. The top individual used 281 billion. Fortune estimated that person alone accounted for $1.4 million in cost. The Information reported on the dashboard, and it was removed two days later. Mark Zuckerberg himself did not crack the top 250. Android Headlines said Amazon had seen similar patterns internally.
The distinction is that Uber’s ranking system came from the company, while the rankings at Meta and Microsoft seem to have sprung up from employees. Nobody needed to tell people to do it. The incentives did the job.
Jay Parikh, head of CoreAI and an executive vice president at Microsoft, spoke to the issue in an internal memo in August. “Tokenmaxxing is not the real goal we should be pursuing. I want everyone focused on outcomes that actually create change for customers and for the business,” he wrote.
The article’s read on that memo was blunt: Microsoft had figured out that a meaningful chunk of employee token spending was paying for signaling, not productivity.
Business Insider also compared the AI spending figures in the spreadsheet with the compensation figures in that same document. At least so far, heavier AI users had not gotten bigger raises, bonuses, or promotions because of that usage.
A much larger AI bill surfaced from inside OpenAI
A day after Microsoft began pulling back Claude Code licenses, a far larger AI bill popped up on social media.
On May 15, 2026, OpenClaw developer Peter Steinberger posted about CodexBar, his menu bar utility, saying a new version showed API costs more clearly. But the screenshot grabbed attention for another reason: $1,305,088.81 in OpenAI API usage over 30 days, 603 billion tokens consumed, 7.6 million requests, and GPT-5.5 as the main model. On the day of the post alone, spending was close to $20,000.
The article said Steinberger joined OpenAI in February 2026, so the bill would plainly be covered by OpenAI. The spending came from about 100 Codex coding agent instances running in parallel, even though only three people were maintaining the OpenClaw open-source project.
Steinberger described the effort as a test of one question: if tokens were free, how would software be written in the future?
At more than $1.3 million, the total was about 46 times the Microsoft employee’s $28,000 record. But the article made a hard distinction between the two situations. In one, about 100 agents were doing work for a three-person team. In the other, a person was pushing up usage to satisfy a performance signal. One bill showed output capacity. The other showed posture.
Microsoft changed course three times in less than a year
Over a longer stretch, Microsoft’s stance on internal AI costs changed fast.
In December 2025, the company opened Claude Code to thousands of employees and encouraged them to reshape workflows with AI.
In May 2026, Microsoft took Claude Code licenses away from most employees, citing “toolchain unification,” while still leaving a route to Claude models through Copilot CLI.
By August 2026, even that route had tightened. CNBC reported that the default model in the employee version of GitHub Copilot had been switched to OpenAI’s GPT-5.6 Sol. GitHub described Sol not as a budget tier, but as the highest reasoning option in the GPT-5.6 family. Coverage elsewhere treated the move as a shift in who pays the bill, not a move to a cheaper tier. The article said reports described Sol as more cost-effective than the Claude family, while Android Headlines said Microsoft was no longer encouraging the use of third-party coding tools such as Claude for programming work.
First Microsoft cut a product. Then it changed the default model. Then it added budgets and dashboards. Control moved from “what tool are you using” to “how much did you spend this month.”
And the timing matters. The May 2026 license cuts were linked to a June 30 deadline, the last day of Microsoft’s fiscal year. Budget controls and monitoring showed up in the first two months of the new fiscal year. Read together, those steps suggest the company ended one fiscal year by shutting down a runaway cost center, then began the next by turning AI into a formal budget line.
The article called this a full shift from “experiment” to “financial item” in about a year. It compared the pattern to cloud computing: first companies told everybody to move to the cloud, then bills ballooned, then they created dedicated cloud cost-management roles. Same logic here. Just applied to tokens.
Three months earlier, GeekPark’s piece "Microsoft Hits Pause on Vibe Coding" argued that the real problem was not that AI had become too expensive, but that organizations had not changed, and most companies would not change their structure quickly. Microsoft’s answer here was not redesigning the organization. It was budget control: updated internal guidance, department caps, and dashboards for individual spending. As Parikh wrote in the memo, Microsoft would manage token spending “with the same discipline we use to manage every other critical resource.”
The article ends on a narrower point about incentives. Microsoft introduced the internal tool that made token usage visible in the first place. Employees then drove the number higher, and the company answered by adding another dashboard to reward thrift. The metric flipped direction. The behavior around visible metrics did not. For Chinese internet companies, the article argues, Microsoft may simply be further down the same road. Some companies in China are still in what it calls the December 2025 phase: handing out quotas, setting examples, and writing AI usage rates into OKRs. Microsoft has already reached the later chapter: encourage usage, rank it, then audit the bill.
The closing argument is pretty narrow and pretty sharp. The hardest thing to manage is not AI pricing itself. It is how people react to visible numbers. As long as there is a column everybody can see, somebody will try to push it higher.
