Apple’s Reported $200-a-Day AI Cap: An Allowance Is Not a Bill
Apple is encouraging customers to ask more of Siri while reportedly asking some of its engineers to spend less on AI. According to Bloomberg’s November 4 reporting, Apple cut the daily AI spending allowance for some engineers from $300 to $200 in October. The number that made me pause was the missing one: how much of that $300 were people using?

A one-third cut to a ceiling can save nothing
Consider two hypothetical workdays. On the first, an engineer spends $40 asking a coding assistant to explain a failure and suggest a test. The bill is $40 under either cap. On the second, a longer investigation would use $250. Under the new ceiling, the last $50 of that plan needs a different approach. Perhaps a smaller model finishes the work; perhaps it takes another day. Those possibilities have very different costs for the team.
TechCrunch’s account of the Bloomberg report also describes possible lower limits, including $100 for certain engineering allowances and a figure as low as $25 outside engineering. Those are reported or contemplated limits with different scopes. There is no basis here for multiplying $200 by Apple’s entire workforce and calling the result its daily AI bill.
Siri and an employee’s coding assistant have different budgets
On the customer side, Apple describes Siri AI as a beta feature that can draw on personal context and work across apps on compatible devices, subject to language and availability requirements. An employee paying for model use during a coding session is a different customer with a different workload. The reported allowance tells us neither the cost of answering a Siri request nor whether an individual Siri feature runs locally.
There is a public number for the broader context. Apple’s fiscal 2026 annual report puts research and development spending at $45.765 billion, up from $34.550 billion. That covers R&D across Apple. It is not a separate disclosure of money spent on tokens, and it cannot tell us what this internal change will save.
The budget I’d put on a small project
For one of my own publishing jobs, I’d set a limit on a finished piece of work: say, extracting a table correctly and checking it against the source. I can try a cheaper model for the extraction, keep the original document handy and reserve a more expensive tool for the rows that don’t reconcile. That’s also where a local model on hardware I already own might be useful. A successful cheap run saves money. Three cheap runs and an afternoon repairing the table might not.
Comments
Post a Comment