The AI bill
Apple reportedly cut some engineers’ daily AI budgets from 300 to 200 dollars, as AI per employee becomes a cost to manage like payroll.
On 4 November Bloomberg reported, as relayed by 9to5Mac and MacRumors, that Apple quietly lowered the daily AI spending cap for some engineers from 300 to 200 dollars in October. It is planning further cuts that could take some engineering teams to 100 dollars a day, a limit many groups outside engineering already have, and it has considered as little as 25 dollars for some teams. Some teams are being encouraged to use cheaper models or models that run on their own devices, and anyone who needs more can ask senior leadership for an exemption.
Bloomberg’s sources also said few people were using their full 300 dollars a day, and it is unclear how much Apple will save. Apple is not retreating from AI either: 9to5Mac notes that its research and development spending rose 32 percent to 45.8 billion dollars in its 2026 financial year, partly on infrastructure and AI. A cap is not a verdict that AI does not work. It is a sign that the spending has grown large enough to manage.
For two years most companies bought AI like software, a seat per person at a fixed price. Agents that read a codebase, run the tests and try again use far more of it, and more of it is now billed by use. That makes AI a cost per employee that moves every day, and it needs the discipline a company already applies to payroll. Here is how, without losing what the spending buys.
Not only Apple
Apple is not alone. In June an Uber spokesperson confirmed to Bloomberg that the company limits every employee to 1'500 dollars a month in token spending per AI coding tool, after it had overrun its AI budget earlier in the year. In October 404 Media reported that Accenture told many employees it would now spend only 29 dollars a month per person on Claude, some of whom had been spending as much as 1'000 dollars a month.
The vendors see the same shift. On Microsoft’s earnings call on 28 October, Satya Nadella said that credit consumption from agentic features on GitHub had more than doubled since usage based billing went into effect. Customers, he said, buy credits and want to use them efficiently: “We don’t want them to overbuy one thing and then feel like they’re token maxing”.
The arithmetic
At Apple’s old cap, an engineer could spend 300 dollars a day. Over a month of about 20 working days that is up to 6'000 dollars, and for a team of ten, 60'000 dollars. Uber’s ceiling with two coding tools in use is 3'000 dollars a month per person; Accenture’s is 29.
These are ceilings, not average spending, but sums of this size need what payroll and cloud bills already have: an owner, a forecast and a monthly review. A salary is fixed, while an AI bill moves with how the work is done: which model runs, how much context it has to read, how often it starts again. A cap limits the spending. It does not make it any better.
Route routine work
The first lever is to stop sending routine work to the most expensive model. At its Universe conference on 28 October, GitHub made Copilot Fusion generally available on all Copilot plans. Instead of picking a model, a developer chooses Fusion and one of three effort levels, and Copilot decides which models handle each prompt. In one of its workflows an efficient model drafts the answer, and a quality check either accepts it or passes the task to a stronger one. Administrators decide which models Fusion may use, and each model is billed at its standard rate.
GitHub says Fusion Max scored within two points of Claude Code with Opus 5.5 Max on the Artificial Analysis Coding Agent Index, at about half the average API cost per task, with about half the tokens and about 40 percent less time. Nadella went further on the call: “It delivers frontier-level quality up to 40 percent faster, and at up to 77 percent lower cost.” Answering analysts, he gave 45 and 75 percent. These are the vendor’s own figures, and “up to” describes the best case; GitHub’s own benchmark puts the saving at about half.
The price gap behind this is large. Anthropic’s Haiku 5.5 costs 0.10 dollars per million input tokens for shorter prompts, a twentieth of Sonnet 5.5. Anthropic says it is “designed for high-volume, cost-sensitive tasks”, and that the larger models remain the better choice for complex agentic coding. The aim is the cheapest model that is good enough for the task, which Apple is now asking some teams to find by hand.
Budgets and shared context
The second lever is where the budget sits. A flat cap per person treats an engineer rebuilding a large codebase like a colleague summarising email. In our view a budget per team works better, with an owner who sees the spending every week and a clear route for exceptions. Apple’s exemption process is the right instinct.
The third lever is context. Much of what an agent reads is the company itself: who the client is, how the company writes an offer, where things are kept, which rules apply. When every agent and every person rebuilds that from scratch, the company pays for it each time, in tokens and in corrections. Context that is written once, kept current and shared costs less to read and leads to fewer wrong turns. Providers price reuse accordingly: on Haiku 5.5, reading context from the cache costs a tenth of the normal input price.
Measure output, not usage
The last lever is what you count. A ranking of tokens used rewards spending rather than results, the token maxing that Nadella said serves nobody’s purpose. In our view the useful measures are the ones a company already understands: changes merged and still working a month later, tickets closed, offers sent, cost per finished task.
A team that spends more and delivers more may be the best investment in the company. A team that spends more and delivers the same is the one to talk to. Managed that way, the bill can shrink without the benefit shrinking with it.
Where we stand
This is how we build at crue. North puts every number that matters for a decision in one view, the real finances included, which is where an AI bill belongs: next to what the work brought in. Twin gives a company’s assistants and agents one brain, with its knowledge, rules and tools connected once, so every answer starts from the company instead of from zero. Both are live.
