Apple has lowered the internal AI usage cap for some engineers from $300 to $200 per day, Bloomberg reported on November 4, 2026. The change took effect in October, and further reductions are reportedly under consideration. However, a usage cap is the maximum an employee is allowed to spend, not the amount actually spent. Cutting a $300 cap to $200 does not necessarily reduce spending by the same proportion. Apple's annual R&D expenses, meanwhile, have risen sharply, so it is worth looking carefully at what the move is meant to achieve and how much cost it might actually save.
Daily cap lowered to $200 for some engineers
According to an excerpt of Bloomberg reporter Mark Gurman's article distributed in Japanese, Apple has cut the AI usage cap for some engineers to $200 per day and plans to reduce it further to $100. In many non-engineering departments, the current cap is reportedly $100, and some teams are considering lowering it to $25. The change implemented in October should be distinguished from the measures that are only being considered.
| Group | Previous daily cap | Change | Status at time of report |
|---|---|---|---|
| Some engineers | $300 | Lowered to $200 | Implemented in October |
| Some engineers | $200 | Further cut to $100 | Planned / under consideration |
| Some non-engineering teams | $100 | Cut to $25 | Under consideration |
Employees who need more AI capacity for their work can reportedly apply to executives for an exemption. While usage spending is being restricted, there is still room to exceed the cap when a need is recognized. However, it has not been disclosed what kinds of work would qualify for an exemption or how much additional usage would be allowed.
An Apple spokesperson reportedly declined to comment, and the timing of the additional reductions has not been announced. The report concerns some employees only; it does not mean the caps for all engineers or all staff have been changed uniformly. Moreover, a change to a daily cap limits what individual employees can spend. It does not mean that the company's overall AI-related spending or R&D budget has been cut by the same proportion.
R&D spending up about 32.5%, a separate matter from the AI limits
Apple's R&D expenses for fiscal 2026 came to $45.765 billion, up about 32.5% from $34.55 billion the previous fiscal year. These figures come from the annual R&D comparison on page 1 of Apple's unaudited consolidated financial statements.
| Fiscal year | Year-end date | Annual R&D expenses (millions of dollars) |
|---|---|---|
| Fiscal 2026 | September 26, 2026 | 45,765 |
| Fiscal 2025 | September 27, 2025 | 34,550 |
The growth rate is (45,765 − 34,550) ÷ 34,550 × 100 ≈ 32.5%. The comparison uses the "Twelve Months Ended" column of the financial statements, not the quarterly R&D figures shown in the same table.
That said, the $45.765 billion is Apple's total R&D spending and includes costs for technologies and products other than AI. The statements do not show how much Apple invested in AI alone, or how much of the increase in R&D spending is AI-related. Conversely, the reduction in employee AI caps is no basis for concluding that Apple is cutting overall R&D spending.
In the first place, annual R&D expenses and employees' daily usage caps differ in both scope and time period. R&D expenses are actual costs recorded in consolidated results, while a daily cap is merely the amount an employee is permitted to use. If employees do not spend up to the cap, the difference is never incurred as actual spending, so the two figures cannot be compared directly to calculate savings.
An increase in R&D spending and a tightening of internal AI allowances are therefore not contradictory. Rather than concluding that Apple is scaling back its AI investment overall, what matters is how much is being spent on which purposes and what results it is producing. Apple itself, however, has not explained that these limits are intended to rebalance budget allocations by use case.
How much AI $200 a day buys depends on how it is used
AI pricing varies widely depending on the model and the type of processing. For example, according to Anthropic's official API pricing page, the base rate for Claude Sonnet 5.5 is $2 per million input tokens and $10 per million output tokens. A token is a unit that measures the amount of data, such as text, that an AI processes, and even for the same model, input and output are priced differently. Feeding in large volumes of material and generating long text or code have different cost structures even at the same token count.
However, these are Anthropic's publicly listed general API rates, not what Apple actually pays. The terms of Apple's contracts with AI service providers, and the breakdown by model it uses, have not been disclosed. It is therefore impossible to convert a $200-per-day allowance into a specific number of tokens or tasks.
AI pricing also includes separate rates for cache usage, apart from standard input and output. Reusing previously processed information can reduce the burden of repeatedly processing the same data in some cases. In other words, how much AI $200 buys depends not only on a model's unit price but also on the ratio of input to output and how caching is used. The cap amount alone does not reveal how much work can actually be completed.
One way to hold down payments to external AI services is to run models on-device. In its developer-facing Core AI documentation, Apple explains that on-device AI processing has no dependence on servers and incurs no per-token charges. If processing can be done without sending requests to external servers, usage-based API charges can be avoided.
Still, Core AI is a reference example for understanding the cost structure of on-device AI processing; it has not been confirmed as a technology Apple introduced in connection with these internal AI limits. Nor does the absence of token fees mean that device costs, power consumption, system management and similar expenses drop to zero. Whether on-device processing can match the quality and speed of high-performance cloud models also needs to be verified for each use case.
Furthermore, lowering AI usage costs does not necessarily lower overall operating costs. If switching to cheaper models or limiting processing capacity degrades the quality of responses, employees may need to spend more time checking, correcting and rerunning tasks. Judging the cost-effectiveness of AI therefore requires considering not only API bills but also the time and effort needed to produce deliverables of the required quality.
How much does lowering the cap actually save?
In the Japanese excerpt of Bloomberg's report, some employees reportedly noted that hardly anyone used up the previous $300 allowance. If an engineer were using only $100 worth of AI per day, lowering the cap from $300 to $200 would not reduce actual costs unless their usage changed. The cut in the cap amount is about 33.3%, but that cannot be taken as the rate of reduction in Apple's AI-related spending.
Calculating actual savings would require knowing the number of employees affected and how much AI each was using before the change. Beyond average usage, it also matters how many employees were spending close to the old cap. The same cap change affects someone who was already spending $200 or less very differently from someone who was spending close to $300. But the number of affected employees, their usage records and the total savings from this measure have not been disclosed, making it difficult to estimate the financial impact accurately from the outside.
There is also not enough information at this point to judge the effect on productivity. Unless it is confirmed whether employees who hit the cap were still able to complete necessary work and maintain the quality of their output, it cannot be said that the cost reduction improved efficiency. The exemption application process for executives could help preserve necessary AI use, but the actual impact will also depend on how many applications are approved.
The report also says that some people inside the company are worried about job cuts. This reflects how employees perceive the situation, however, and does not show that the AI usage limits are tied to any specific plan to reduce headcount. No measurements have been disclosed showing that the cap changes raised or lowered productivity.
Whether the measure amounts to sound cost management cannot be judged without comparing actual usage with work results. If spending can be reduced without preventing necessary work from being completed or worsening quality and rework, revisiting daily caps could be an effective way to manage costs while sustaining AI use. If, on the other hand, the limits increase work time or the burden of corrections, a superficial reduction in AI fees will not necessarily lead to lower costs for the company as a whole. The focus going forward will be how much spending Apple actually cuts through this measure, and whether it can maintain productivity on the development floor.