The number inside Apple that best describes the past two years of corporate enthusiasm for artificial intelligence is $300. That was the daily allowance some engineers were given to spend on model calls, enough to keep an agent working through a long day without anyone asking what it cost.
Bloomberg reported on November 5 that the company has been pulling that number down. Apple cut the cap for some engineers from $300 to $200 last month and now plans to reduce it further for some engineering teams, with the floor as low as $100 a day. Some non-engineering groups already operate at $100. The company is considering taking some teams to $25.
Employees who need more can apply to senior management for an exemption, according to the report. Apple did not respond to a request for comment.
The reversal is abrupt in a company that spent the past year urging staff to use generative tools as much as possible. What changed is not the ambition but the bill. Model usage priced per token behaves differently from software licensed per seat: a fixed number of licenses produces a predictable cost, while an open-ended allowance produces a cost that scales with how inventively employees use it. Multiply a few thousand engineers by a few hundred dollars a day and the annual figure becomes a budget line that finance departments notice.
The per-day cap is a crude instrument, and that is partly the point. It is easy to administer, easy to explain, and it forces teams to make choices rather than treating inference as free. A team that hits the ceiling has to decide whether a task is worth an exemption request, which is a way of discovering which uses of the technology actually pay for themselves.
Apple’s staff have access to models from Anthropic, OpenAI and Google, a menu that reflects how quickly the market for frontier models has become a procurement decision rather than a research one. The report says some teams have been asked to move to cheaper models or to models that run locally on the device. That instruction carries an obvious logic for a company that designs its own silicon: work that can be done on hardware Apple already sells is work that does not generate a bill from a supplier.
The cost backdrop is not one of retrenchment. Apple’s research and development spending rose 32% in fiscal 2026, and total operating expenses climbed 22% to $75.5 billion. Those are the numbers of a company investing heavily, not one in retreat. The constraint being imposed sits at the margin of a specific kind of spending, the kind that arrives as a metered utility rather than as a salary or a building.
None of this is exotic for a company of that size. What is unusual is the speed. A budget of $75.5 billion does not grow 22% in a year without something new inside it, and model usage is the newest category most large engineering organizations have bought since cloud computing. The accounting for it is still unsettled: whether a model call counts as research, as infrastructure or as a cost of goods sold depends on which team is making it, and that ambiguity is one reason quotas appeal to finance staff.
There is a strategic question underneath the accounting. Apple has argued that its advantage in AI comes from running models on its own devices, where the marginal cost of a query is borne by the owner of the phone rather than by the company. A budget that pushes engineers toward local models is consistent with that view. It also concentrates the company’s internal experience on the constraints of on-device inference: memory, battery, latency and the quality ceiling of a model small enough to fit in a pocket.
The alternative reading is less flattering. If Apple’s own models were strong enough for the work its engineers do, the cheaper path would also be the obvious one, and no quota would be needed. A cap that pushes teams toward third-party models flagged as low-cost suggests a company managing a dependency rather than exercising a preference.
Either way, the quota changes behavior inside the building. Engineers who previously chose a model by capability now choose by price, and the tools that win internal adoption will be the ones that deliver acceptable output at the lowest metered cost. That is how enterprise software markets have always worked, and it is a useful test for model providers who have spent two years arguing that quality alone would carry the day.
The comparison with Apple’s own history is instructive. The company has never been shy about spending on components it considers essential, and it has been equally willing to squeeze suppliers of parts it regards as commodities. The question the new caps raise is which category frontier models now occupy in Cupertino’s internal ledger.
For employees, the practical effect is a working day with a meter attached. For the model providers, it is a signal that their largest customers are beginning to treat inference as a cost center to be optimized rather than a capability to be maximized. For Apple, it is a bet that the price of intelligence will keep falling fast enough that a $100 daily budget buys next year what $300 buys today.
Whether that bet is right will show up in what the company ships, and in whether the engineers who lost their allowance produce the same work with cheaper tools. The quota is easy to reverse. The habit of asking what a query costs is not.
