Cost Fell. Did AI Actually Work?

Ramp just put out their monthly AI spending report. For the first time since they started tracking it, the top 1% of AI spenders spent less on AI last month than the month before.

I like using Ramp’s data for these updates because it’s real spend, not a survey. Companies pay for AI through Ramp cards and bill pay, so this is what businesses actually paid, not what they say they’re planning to pay. The top 1% of spenders went from about $8,000 per employee in July down to $7,200 in August.

Cost is an input when you’re measuring whether AI is working for your firm or your clients. It is one piece of the scorecard, not the whole thing. Cost comes down to price per token times how many tokens people actually use, and price per token has been falling fast across every major provider. So when total spend drops, you genuinely cannot tell from that number alone whether usage went down, price went down, or both.

The effective price per million tokens is down 41% since its March peak, sitting around 67 cents now. I’d add one caveat that Ramp’s report doesn’t spell out directly. Anthropic and OpenAI are both racing toward IPOs, Anthropic reportedly this year and OpenAI likely next, and they’re competing hard on nearly everything, including price. We saw a version of this play out just this week. OpenAI reportedly spent something like $23 million in a month trying to be first to solve a well known unsolved math problem, seemingly to beat Anthropic to the punch. When two companies are spending at that level to win bragging rights, it’s a reasonable bet they’re also willing to eat margin on token pricing to keep businesses from switching providers. Some of this price drop is probably genuine efficiency, and some of it is probably short term competitive spending that won’t hold once one provider pulls ahead.

Frontier models, the most capable and most expensive option from each provider, peaked at over half of all tokens used back in August and have been losing share since, even with new frontier releases coming out. That lines up with what we covered last month. The best model is rarely the model most people actually need for day to day work, and more companies are building that into their defaults.

On adoption, Anthropic is holding around 44% of the market and OpenAI is close behind at about 40%, with potential overlap between the two. A lot of businesses are running both, the same way large companies often run AWS and GCP together instead of betting everything on one cloud provider. Having two vendors gives you leverage and some protection if one provider’s pricing or performance shifts.

Every cycle in business seems to pick up a metric that gets more trust than it deserves. Railroads got paid by the track mile in the 1800s, and some of them laid duplicate, curvy track just to rack up miles. The dot com era ran on website eyeballs, and plenty of high traffic sites never turned a profit. The 2010s chased top line revenue and mostly ignored what it cost to get there. Token price and token spend risk becoming 2026’s version of that same pattern if we treat either one as the full picture instead of one piece of it.

About a year ago we put together a lesson on measuring AI the way tech teams do, using four factors: speed, quality and control, experience, and cost. The rule is to always pair two of them together. Did things get faster, and what happened to quality control while that happened? Did cost go down, and what happened to the experience of the people actually using the tool? Looking at cost by itself tells you almost nothing.

Key Takeaways

• Don’t forecast off today’s $0.67 per million token price. Part of that drop is real efficiency, and part of it may be competitive spending ahead of two major IPOs. Track your own effective price per million tokens so you have a number specific to your firm.
• Falling cost does not mean AI is working. Confirm it against speed, quality and control, and experience before you call it a win.
• When a cost line moves, ask what actually changed: more usage, a price drop, or a vendor eating the cost. Each one points to a different next step.
• Keep budgeting AI by vendor and by model, and keep the option to switch between providers open. The competition between Anthropic and OpenAI is working in your favor right now.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: Cost Fell. Did AI Actually Work?