In the week ending October 23, 2026, businesses tracked in the Ramp AI Index pushed 37.1% more tokens through AI models than in the same week a month earlier. Their token bill rose 8.9%. Before that, across August and September, token spend grew just 1.5% in total, after a 42% month-over-month jump in July, according to Ramp data reported by TechCrunch on November 6. Usage is climbing and the invoice is flattening, and one of the reasons is that finance has started saying no.
The most visible no came from Apple. As MacRumors reported from a Bloomberg story on November 4, Apple cut the daily AI allowance for some engineers from $300 to $200 in October, may take some engineering teams down to $100, and has considered dropping some non-engineering groups from $100 a day to as little as $25. Put the two stories together and you get the shape of the next phase of enterprise AI: tokens are cheaper, usage keeps rising, and leaders are reaching for caps. The question this post asks is whether a cap is AI cost management, or just a smaller bill with the same blind spots.
What Ramp’s data actually shows
Ramp builds its index from corporate card and bill-pay transactions across more than 85,000 U.S. businesses, and its token metrics come from the subset of customers using its token spend management product. The tokenomics table tells most of the story. Between the week ending September 25 and the week ending October 23, the blended price per million tokens across OpenAI and Anthropic fell from $0.76 to $0.56, a 26.8% drop. OpenAI’s blended price fell 42.4% and Anthropic’s 15.7%.
Ramp’s lead economist, Ara Kharazian, told TechCrunch the slowdown is not a demand problem. He points to price cuts, including OpenAI halving API prices for GPT-6 Sol and GPT-6 Luna in September and Anthropic’s own reductions, which the company said in a filing cut its revenue growth by about $600 million. He also points to buyers getting more disciplined. “As you get really advanced in using AI, you get better at figuring out how to control it,” he said, describing defaults that keep expensive models off long-running tasks, budgets imposed on employees, and routers that shift work from standard and light models to even cheaper ones. His summary is the line every CFO should keep: “You can use more AI, but use it for less.”
The spread between companies is the other headline. In September, the median firm paying for AI spent $15.70 per employee per month, the top 10% spent $813.43, and the top 1% spent $9,567.53, according to the index. Kharazian told TechCrunch that this top 1% stuttered slightly in August and September, and that the slowdown in that cohort alone was enough to flatten overall spend. The companies that went furthest with AI are the first to tighten. It is the next chapter after your AI got cheaper but your bill did not: prices kept falling, and the heaviest users started managing consumption.
Apple’s $200-a-day answer
Apple’s version of discipline is a per-person daily ceiling. Per the Bloomberg reporting summarized by MacRumors, the budgets cover tokens across services such as Claude, ChatGPT and Gemini; engineers carry the largest allowances because coding burns through tokens; some teams are being steered toward cheaper models or on-device AI; and employees who need more can request an exception from executives. Some Apple employees told Bloomberg that few people were using their full $300 daily limit, so the savings may be small.
That last detail is the important one. If most engineers never touched the old ceiling, lowering it from $300 to $200 mostly changes the outcome for the handful of heavy users, and Apple’s spending policy now treats those people the same way regardless of what their usage produced. A heavy user shipping more reviewed, merged code with agents looks identical to a heavy user looping an expensive model on a task that never lands. It is the mirror image of the leaderboard problem we covered in tokenmaxxing: one rewarded consumption, the other punishes it, and neither measures value.
Apple is not alone in reacting after the fact. We wrote earlier this year about how Uber blew through a year of AI budget in four months and built a cap afterward. Reuters, summarizing the same Bloomberg story, noted that Goldman Sachs restricted access to some AI coding tools after a review of productivity and risk management. The pattern: encourage adoption, discover the bill, then clamp down with the coarsest instrument available.
Why a spending cap is not AI cost management
A cap answers one question: how much are we willing to lose if nothing else is working? That is a legitimate guardrail. But AI cost management has to answer three harder questions that a ceiling cannot. Who is spending, on which provider and model, for which project? What did that spend produce, measured in an outcome the business already tracks? And where is money leaking into idle seats, oversized models, or work that a cheaper route would have handled just as well?
Ramp’s numbers show why those questions matter more now than in July. When blended token prices fall 26.8% in a month, total spend stops being a useful signal on its own. A flat bill can hide a team that doubled its useful output and another that doubled its waste. The CFO’s real metric is cost per outcome, the AI equivalent of unit economics, and that requires joining spend to usage to results. That is the unit-cost gap Deloitte found when it reported that most enterprises can see the AI bill while far fewer know the unit cost.
This is the signal Olakai was built for. Olakai is a vendor-neutral Enterprise AI Intelligence Platform that sits across coding, assistive, and agentic tools, so finance sees spend, adoption, and impact on one record instead of in each provider’s console. Its AI spend governance budgets work as overlapping lenses over the same dollars, by program, provider, developer, department, and project, with threshold alerts at 50%, 80%, or 100% and forecast alerts that fire when the run rate is on track to exceed a limit before the month ends. The point is seeing the trajectory early enough to ask whether the spend is earning its keep.
What AI cost management looks like when it works
Start with attribution. Spend that cannot be tied to a person, team, or project cannot be judged, so the first job of a measurement layer is to connect provider invoices to the people and workflows generating them. Then segment adoption. Our adoption cohorts sort users into Power, Casual, New, and Idle groups, which turns a vague complaint about licenses into a list of seats to reclaim and heavy users to study rather than throttle. Engineering leaders can go further with Olakai Agentic, which compares AI-assisted pull requests with the rest on cycle time, so the $200-a-day question becomes whether a developer’s AI usage shows up in shipped work.
Next, route on purpose. Kharazian’s advanced buyers are already sending routine work to cheaper models, and the savings are real, but routing without measurement is guesswork. Our explainer on model routing covers the tradeoffs; the measurement layer tells you whether quality held after the switch. Finally, tie it to return. A CFO should be able to pull a cost-per-outcome view next to the budget, which is what analytics and custom KPIs are for and what our AI ROI framework is built around.
What CFOs and CIOs should do this quarter
If your company is about to follow Apple, run these steps before the cap goes in, not after:
- Pull a per-person and per-project spend distribution for the last 90 days. If, like Apple, most users sit far below the proposed ceiling, the cap will only bind a small group, and that group deserves an outcome review before a blanket cut.
- Compare cost per outcome, not cost per employee, for your heaviest users. Keep the ones whose usage maps to shipped work or closed tickets, and fix routing or model defaults for the ones whose usage does not.
- Set alerts on trajectory as well as totals, so finance hears about a runaway month in week two instead of on the invoice.
- Reprice your budget assumptions. With blended token prices down 26.8% in a single month on Ramp’s data, last quarter’s per-seat math is already stale.
For the finance and FinOps team, the governance point is that a cap is a policy, and policies need owners, exception paths, and evidence. Apple routes exceptions through senior executives, which scales poorly. Pair every limit with an AI governance rule that says who can raise it, on what evidence, and how the decision gets logged.
Where to start
Pick your three most expensive AI tools and answer one question for each: if you halved its budget tomorrow, which outcome would suffer, and by how much? If nobody can answer, a cap is the only lever you have, and you will end up where Apple and Uber did, cutting first and measuring later. Olakai gives you the other option: a system of record for the entire AI stack that shows who is spending, what it is producing, and where the money is better spent, so the next budget conversation starts with evidence.
Talk to an expert about measuring AI spend against outcomes across every provider, team, and project before your next budget cut is a blunt one.