This is the third beat of a story we have followed since October. Microsoft showed how it cut its own Claude bill by about a third. Accenture then capped the Claude spend it funds centrally for consultants at $29 a month. This week the caps reached two of the most AI-forward employers anywhere, and for the first time they are being drawn by job title. Apple is reportedly cutting the daily AI allowance for its engineers, with lower limits for everyone else. Goldman Sachs has reportedly taken Claude Code and GitHub's coding tools away from most traders and bankers and kept them for engineers.
Neither firm is short of money or conviction. Apple's R&D spend rose 32% to $45.8bn in its 2026 fiscal year, driven partly by AI and infrastructure costs, according to 9to5Mac. Goldman's chief information officer was praising traders who vibe-coded their own tools as recently as July, FStech reports. If these two are deciding who gets what and how much, the per-head AI budget is now a management decision at every firm with a seat licence, whether that is a consultancy, a contractor or a client-side PMO.
Apple: $300 to $200, then $100, then maybe $25
Bloomberg reported on 4 November that Apple quietly lowered the daily spending cap for some engineers using AI models from $300 to $200 in October, and is planning further cuts for some engineering teams to as little as $100 a day. Many groups outside engineering already sit at $100, and Apple has considered taking some of those to $25. Staff who need more can seek exemptions from senior leadership. TechCrunch says Apple has not commented.
Two details matter more than the headline figures. Bloomberg says Apple staff have access to Claude, ChatGPT and Gemini, and some teams are being pushed towards cheaper models or models running locally on their machines. And its sources said few workers were using their full $300 allowance, so it is unclear how much the cuts, part of CEO John Ternus's wider cost programme, will save.
The direction of travel is striking for a company whose then chief executive, Tim Cook, told an all-hands meeting last year that "All of us are using AI in a significant way already, and we must use it as a company as well," as 9to5Mac recalls. That is the awkward truth about per-head caps: if most people never touched the limit, the cap lowers the ceiling only for the minority using the tool hardest, and those are the people a delivery firm should want to study.
Goldman: coding tools for 12,000 engineers, nobody else
The Goldman change is about entitlement rather than budget. According to internal guidance updated in October and seen by Bloomberg, reported by FStech on 4 November, the bank had let staff across the business use Anthropic's Claude Code and Microsoft's GitHub tools for several months. Traders and bankers used them to monitor client flows, model market moves and produce written and visual material. Access is now limited to engineers, a workforce of roughly 12,000, plus the 'strats' whose jobs already combine trading, quant research and software.
After a period of experimentation where we assess a variety of factors, including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term.
The approach is straightforward. Give everyone access for a fixed trial period, assess usage against two specific criteria, and then decide who gets continued access. This is a shift from what CIO Marco Argenti was saying in July, when he highlighted how engineers could improve applications built by colleagues with no technical background.
Why does the bill behave like this?
Digital Storm's 3 November issue runs under the headline "Your 40th message costs 14 times your first", because Claude re-reads the whole conversation before every reply; the workings sit behind a sign-up wall, so treat the figure as the author's. Its 5 November piece, paywalled after the introduction, cites SemiAnalysis pricing each major plan at API list rates: on mid-tier models Claude plans deliver roughly 5.4 to 5.6 times the value of comparable ChatGPT plans, and one $200 Max seat buys about $11,726 of Opus 5.5 usage a month.
That is a subsidy set by policy, and OpenAI has already halved the allowance on its own $200 plan. A seat is cheap, a long agent session is expensive, and the vendor can move the line whenever it likes. Last week we noted Amazon leaning on staff to stop 'tokenmaxxing' for its own sake. Apple and Goldman have the same instinct for arriving at the top of the market.
What a delivery firm should set
Four decisions fit on one page for a board.
Decide entitlement by role before budget by head: a QS, a planner and a bid writer need a general assistant with document access, while a coding or agent tool belongs with the digital team and anyone who can show a repeatable use.
Set a daily or monthly ceiling per role high enough to be rarely hit, then watch who hits it, because Bloomberg's reporting suggests the people at the limit are the ones doing the most with the tool.
Give every cap an exemption route a line manager can approve in a day, as Apple reportedly has through senior leadership, so a genuine win is not stuck in a form while the deadline passes.
Measure value in outputs a client recognises, such as hours saved on a cost report or tender comparison, because a token count says nothing about whether the work was better.
Takeaway
Caps spreading from vendors and consultancies to Apple and Goldman settle one argument: per-head AI budgeting is now normal practice. The risk is in how it is done. A cap set at the median user starves the people at the edge, and the edge is where the useful discoveries happen. Set entitlement by role, set the ceiling high, build a fast exemption route and measure what came out. Keep a list of who asked for more and why, because that list is your next business case.
If your firm is deciding who gets which AI tool and how much they may spend, the Project Flux newsletter tracks every cap, seat and pricing change as it lands, with the delivery view attached. Subscribe and send us your policy; we would like to compare notes.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
