Apple Cut Engineer Token Budgets From $300 to $200 a Day. A Cap Is the Wrong Number
Apple quietly lowered daily AI spending limits for engineers and is studying deeper cuts, while Nvidia argues the opposite. For anyone producing images and video with models, a daily dollar cap measures the wrong thing. Cost per accepted asset is the number that survives contact with real work.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

Last month, Apple quietly lowered the daily AI token budget for some of its engineers from $300 to $200. According to Bloomberg's reporting, the company is now studying a further cut to $100 for certain teams, and some non-engineering groups would drop from $100 to $25 a day. Anyone who needs more can request an exception, which will be reviewed. Apple staff have access to Claude, ChatGPT, and Gemini, and some teams have reportedly been told to try cheaper models or ones that run on the device.
The detail that makes the story useful is not the cap. It is that, according to employees quoted in the reporting, few people were using the full $300 in the first place.
A budget nobody hits is not a cost control. It is a measurement of how the tool actually gets used, presented as a decision about spend.
The two positions, stated honestly
Apple is the largest company in the market by capitalization and spent $45.765 billion on research and development in its last fiscal year, up 32% year over year, with part of the increase attributed to AI. If Apple is trimming per-person AI spend, the obvious question is who can afford not to.
Nvidia answers it from the other side. Jensen Huang said on the All-In podcast that he would be worried if an engineer earning $500,000 spent less than $250,000 a year on tokens. Asked whether Nvidia spends around $2 billion a year on tokens for its engineers, he said they are trying to.
Both positions are coherent once you notice what each company sells. Nvidia sells the chips the models run on, so every token burned is demand for compute. Apple pays the bill and ships the product. The advice you get about AI spending tends to track which side of that invoice you sit on.
We have run this experiment before, badly
Apple is late to a correction that started earlier in the year. Microsoft found a single employee who had burned $28,000 in tokens in one month, and vice president Jay Parikh wrote internally that "tokenmaxxing is not what we are optimizing for". Amazon deleted an internal leaderboard that ranked engineers by token consumption in May, then published guidance discouraging the practice without imposing hard limits. Meta had its own internal race over who consumed the most.
Then there is Revolut, which took the structural route instead of the budgetary one. It moved roughly 2,000 engineers to GLM, an open-weights model it can run on its own servers, and reports paying about a tenth of frontier-model pricing with a fixed bill rather than a per-request one. Before the switch, $300 a month per engineer for Claude Code did not last the month.
Four companies, four different instruments: a per-person cap, a memo, a deleted leaderboard, and a change of supplier. Only one of them changed the cost structure rather than the spending ceiling.
Why a daily dollar cap is a bad instrument for visual work
Code has a convenient property for budgeting: a task either compiles and passes or it does not, and the cost per finished change is roughly measurable. Image and video work does not behave that way. The unit you buy is an attempt, and the unit you sell is an accepted asset. Between them sits a rejection rate that depends on the model, the prompt, the reference, and the taste of whoever is reviewing.
That gap is where daily caps go wrong.
A $200 day on a model that needs twenty attempts per keeper buys ten usable images. The same $200 on a model that needs four attempts per keeper buys fifty. Neither number appears in the budget. The cap is identical, the output is not, and the team that looks expensive per token may be the cheapest per delivered frame.
Video makes the distortion worse, because cost scales with duration and resolution rather than with a flat per-call price. A daily cap hit at 11am on a single 8-second shot tells you nothing about whether that shot was worth doing.
What to measure instead
Cost per accepted asset, not per token
Take the last project, tag every billed generation with the model or workflow that produced it, and divide that model's total cost by the number of its assets that survived review. Compare models on similar assets — a character sheet against a character sheet, a final frame against a final frame — because asset types do not cost the same to produce. That number is comparable across models, across vendors, and across weeks. It is also the only number that maps to a client invoice.
The p90, not the ceiling
If a typical day costs $40 and one in ten costs $180, a $200 cap is not a control, it is a speed bump on the days that produce the most work. Set the limit above the ninetieth percentile of real usage, and put the review on the exceptions instead of on everyone.
Route by stage
Prompt drafting, variation grids, and composition tests do not need the frontier model. The final frame, the final seconds of motion, and anything with a face or a product in it usually do. Most of the wasted spend in a creative pipeline is not on the expensive call, it is on using the expensive model for the cheap decision. The cost-lane argument for video models makes the same case with per-shot numbers.
Log acceptance rate next to spend
Cost per accepted asset is the price per attempt divided by the acceptance rate: $1 per attempt at a 10% acceptance rate is $10 per keeper, and $10 per attempt at 60% is about $16.67. A cheap model with a 10% acceptance rate can still win that comparison, but neither the sticker price nor the acceptance rate alone tells you, and nobody notices until someone writes all the numbers in the same row. This is the metric most teams skip, and it is the one that decides whether a budget cut actually saves money.
The part Apple got right
A cap with a documented exception path turns spending into a conversation about value instead of a wall. Someone who can justify $400 a day for a week gets it, and the justification is the useful artifact: it says which work is worth frontier pricing. The failure mode is a cap with no path, where teams quietly switch to worse models to stay under a number nobody chose for their workload.
If you are setting this up for a studio or a solo practice, the practical sequence is short. Measure cost per accepted asset for two weeks. Route the cheap decisions to cheap models. Set the cap at p90 of what the good work actually costs. Then leave it alone long enough to see whether the output changed, because the only budget that matters is the one that keeps producing frames someone will pay for.