Uber now caps employee spending on AI coding tools at $1,500 per month, per tool after burning through its entire 2026 AI budget in four months. Microsoft is racing to replace expensive outside models with its own.
The pattern is clear enough that companies are scrambling to figure out how to budget for AI.
Uber has started capping how much its employees can spend on AI coding tools such as Claude Code and Cursor, setting a limit of $1,500 per month per tool. The move follows Uber burning through its entire 2026 AI budget in about four months because they get billed by the token, a small chunk of AI-generated text, rather than a flat monthly usage license.
AI expert Simon Willison did the math: An engineer using two tools at the cap would run about $36,000 a year, roughly 11% of a typical Uber software engineer's $330,000 total pay. He called the cap a rational response, and noted it actually proves how much value the tools deliver, since people were willing to spend that much.
The same squeeze is hitting Microsoft. AI chief Mustafa Suleyman told Bloomberg that "Anthropic is extremely expensive" and that many people are urgently looking for alternatives, with Microsoft's goal being to reduce and ultimately eliminate that cost. Microsoft answered by launching seven of its own AI models, including one at a lower price that it says matches Anthropic's Claude Opus 4.6 on a coding test. Others are building workarounds: The AI coding company Factory launched a feature that automatically picks the cheapest capable model for each task, claiming it cuts costs 25% without losing performance.
On Episode 218 of The Artificial Intelligence Show, SmarterX founder and CEO Paul Roetzer examines what the spending crackdown means.
$1,500 - Uber's new monthly per-tool spending cap on AI coding tools used by its engineers
4 months - How fast Uber burned its entire 2026 AI budget
25% - Percentage of costs the company Factory claims it can save users by picking the best, most cost-effective AI tool
100 billion - Tokens OpenAI's top spender uses in a month
210 billion - Tokens one OpenAI employee reportedly spent in a single week
AI agents are being used heavily now. As OpenAI CEO Sam Altman put it, the issue surfaced suddenly at the start of 2026. "The issue never came up. People were totally happy with the amount they were spending," Altman said, before adding that AI costs are now "a huge issue." Roetzer notes this began being discussed first on platforms such as X, where the heaviest AI users gather, often months before mainstream coverage caught on. "I think X is actually a really good signal for what's going to be really important," says Roetzer.
AI token pricing does not correspond to value. The deeper problem is that paying by the token assumes usage equals value, and it does not. Sierra, the company run by OpenAI board chairman Brett Taylor, published a piece arguing for outcome-based pricing instead, where customers pay for results rather than usage. But Sierra cautions it is not simple. As the piece put it, "outcome-based pricing is more complex than seat-based or consumption pricing. People telling you it's simple are selling something." It only works where the AI is highly autonomous and its results can be cleanly attributed to it.
The smartest move is a token audit by role. Roetzer's practical fix is to stop handing out blanket access and instead map AI use to actual jobs. "You almost need to do this token analysis by roles and workflows," says Roetzer. Most tasks a salesperson does, for example, do not need the most expensive model. He also says the real wildcard is AI literacy because the same budget produces wildly different results depending on a user's skill. The average knowledge worker has no idea what the difference is between a model's various settings, so a basic email can get written by the most expensive option when a cheaper one would do.
"The percentage of budget means nothing if the person doesn't know how to use the tokens wisely. Simple AI literacy would dramatically change the token burn within companies."
— Paul Roetzer, founder and CEO of SmarterX, Episode 218 of The Artificial Intelligence Show
Roetzer's framing for SmarterX is a very useful one for leaders to borrow. Rather than starting from a token budget, he starts from headcount. Look at the roles you would have hired over the next year, decide which ones AI now lets you skip, and redirect a slice of that saved salary budget into tokens for the people who remain. The focus should be revenue per employee, not a flat percentage of salary spent on AI. That's because a percentage rule rewards heavy users regardless of whether they create value.
The harder layer is agentic AI, meaning AI that takes multiple steps on its own without a human involved. A single research request can quietly spin up several agents running in parallel, each burning as many tokens as a full chat session. Even experienced users struggle to predict that spend, and administrators often cannot control it for their teams. This is why the answer cannot be manual oversight. It has to be some combination of automatic routing, better literacy, and new governance tooling because no manager can police token use by task.
Model routing becoming a standard term and practice. Coinbase CEO Brian Armstrong reported the company has kept AI costs roughly flat even as usage grows exponentially, largely by routing prompts to cheaper models when the top-tier one is not needed. Expect routing, whether built by the labs or by third parties, to become the default way companies control spend.
Falling model costs solve the problem. The cost of any given model tends to drop sharply within a year, which means much of today's spending crisis could ease on its own. The companies that win will be the ones that learn to match the right model to the right task instead of throwing the most expensive model at everything.
Only 25% of organizations have reached the Scaling phase of AI adoption, according to the 2026 State of AI for Business Report, while 47% are still running limited pilots. That gap is exactly what makes runaway AI bills feel so painful: Most companies are paying for experiments, not a scaled capability, so the spending climbs while the return stays fuzzy.
The report is built on more than 2,100 responses from professionals across roles, functions, and industries, and it maps where organizations actually stand on adoption, governance, training, and tooling. If your AI costs are climbing faster than your results, it is the clearest benchmark for figuring out whether you are building real capability or just funding pilots. Read the full report →