Bloomberg says Apple cut some engineers' AI budget from $300 to $200

Bloomberg reports that Apple cut the daily AI spending cap for some of its own engineers. The new cap is $200 a day. The old cap was $300. The cap sets what one engineer may spend on model API calls in a day. Hacker News surfaced the report through the newsletter Stork AI Daily.

The Bloomberg report carries three facts. The new cap is $200 a day. The previous cap was $300. The cut applies to some engineers, not every one of them. The Bloomberg article sits behind a subscription wall. This item does not rely on any detail beyond those three facts.

Why a per-head daily cap is a new thing

A year ago, the companies who bought the most AI code work treated it as a flat software line. A team bought a block of seats, and the cost was known in advance. The move to agent-driven coding changed the shape. A single long agent run can spend ten or a hundred times what a human coder would. A Claude Code session that compiles, tests and retries a hard refactor can cost hundreds of dollars on its own. At a thousand engineers, that is a budget nobody built for last year.

The story here is that the cap exists at all. A large company now meters model spend the way it meters cloud compute: per person, per day, with a number the engineer can see.

What a $200 daily cap looks like in practice

The cap is a ceiling on the worst day, not a target for the average day. A developer who writes code and uses an agent for a few tasks spends well under it. The cap bites on the long agent run. An agent can loop on a failing test for an hour. It can ask the top model to think longer at each step. That is also the pattern a vendor earns the most from. The cap pushes back on the behaviour a vendor's own documentation recommends.

The cap also forces a model choice. A medium model costs a fraction of the best one. An engineer under a cap learns to route the easy work to the cheaper model. The expensive calls stay for the step that needs them. The planner-executor split that open projects have published this year is the same idea.

What the story does not say

Bloomberg did not report how many Apple engineers hold which cap. It did not report the previous average spend before the cut. It did not report how a cap is enforced in practice. It did not report which teams are inside the capped group. A buyer of AI tools has three numbers from the report and no mechanism behind them.

Enterprise teams who set their own per-head caps this quarter will not be copying Apple. They will be copying a public number that an engineering organisation their size decided was the right ceiling.

Source

• Bloomberg (subscription required)
• Stork AI Daily
• Hacker News, item 50421426