$300 a day per engineer. That was the AI token budget some Apple engineers had until October, according to Bloomberg. It is now $200, and more cuts are planned: down to $100 a day, and as low as $25 for some non-technical teams.

Put the old figure in context. At $300 a day across roughly 21 working days, one engineer could spend about $6,300 a month on AI tokens, more than 300 times the price of a $20 consumer subscription. Even the new $100 cap is about $2,100 a month. Multiply that across thousands of engineers and the line item becomes visible at board level, which is exactly where it seems to have landed.

What is being capped

Apple employees can use Anthropic's Claude, OpenAI's ChatGPT and Google's Gemini at work. Access is not being removed; the amount they can spend is. Employees who need more can request an exception from management, and some teams are expected to move to cheaper models or to AI that runs locally on devices.

The cuts fit the direction set by new CEO John Ternus, who has been pushing for a leaner Apple. The company's R&D spending rose 32% in fiscal 2026, driven in part by infrastructure and AI costs.

Reaction inside Apple is mixed, per Bloomberg. Some employees think the old limits were too generous. Others fear the cuts are a prelude to layoffs.

Apple is not alone

Microsoft, Tesla, Uber and Meta have all recently capped internal AI spending, The Decoder notes. For the past year, the mood in big tech was that more tokens meant more output, and some companies even tied usage dashboards to performance reviews. The question has now flipped to output per dollar.

The lesson for smaller teams

Most small companies are nowhere near Apple's numbers; a team on $20 seat subscriptions has a predictable bill. The risk sits elsewhere: API-based agents and coding assistants running in automated loops. One agent stuck in a bad loop overnight can run up hundreds of dollars before anyone notices.

Apple's playbook scales down well. Set a daily cap per person or per team, keep a simple exception path for people who genuinely need more, and match the model to the task. That last step usually saves the most: classification and summarization rarely need a flagship model, and routing them to a smaller one can cut the bill without anyone noticing a difference in quality.

Sources: The Decoder (citing Bloomberg), TechCrunch