In April, the Ramp AI Index counted 34.4% of the businesses in its data paying Anthropic and 32.3% paying OpenAI. It’s the first time Anthropic has led. A two-point gap is a small thing to build a vendor strategy on, and a payment count measures billing, not work. Before any leaderboard touches your renewal, total your own AI invoices for the last 90 days.
Paid AI adoption across all firms sat at 50.6%. Half of the firms in Ramp’s data now pay someone for AI, so the interesting question has moved from whether to buy to what each purchase returns.
Key points
- What does the Ramp AI Index actually count? The Ramp AI Index counts businesses that pay each vendor, using corporate card and invoice payments that run through Ramp’s own platform.
- Why isn’t spend the same as market share? Spend counts invoices, not usage. It can miss companies that buy through enterprise agreements or resellers, and it treats a trial seat like a company-wide rollout.
- Is the price per task rising while the lead grows? Possibly, and Ramp says so itself. Its write-up flags frequent outages, rate limits and growing dissatisfaction with results, and says a recent Anthropic model update would triple token costs for any prompt that includes an image.
- How should you read your own AI bill? Pull 90 days of AI charges from the company card and the accounting system, then tag each by vendor and by team.
What does the Ramp AI Index actually count?
The Ramp AI Index counts businesses that pay each vendor, using corporate card and invoice payments that run through Ramp’s own platform. A firm that bought one seat last spring and a firm running ten agents all day each count once. The index measures who gets billed, not who gets work done.
Ramp wrote that Anthropic’s adoption rose 3.8% in April while OpenAI’s fell 2.9%, and that over twelve months Anthropic quadrupled business adoption while OpenAI grew 0.3%. The monthly moves are small. The twelve-month gap is the data point with weight, and it describes who signs up, which is a different thing from how much gets done.
Look at who’s in the sample too. These are Ramp customers, which means companies with a modern spend-management setup. That’s our reading, because Ramp’s page states no sample size. It also doesn’t say whether the 3.8% and 2.9% are percentage points, so we’ve repeated them as written.
Why isn’t spend the same as market share?
Spend counts invoices, not usage. It can miss companies that buy through enterprise agreements or resellers, and it treats a trial seat like a company-wide rollout. That’s our reading of a payment-based method; Ramp’s page doesn’t say how it handles either case.
Picture what one payment hides. A firm pays for a chat assistant for 40 people and a coding tool for four developers, and Ramp records two vendors. The four developers may generate most of the usage and most of the cost. A 60-person company needs its own cost per accepted piece of work, and no index publishes that.
A counter-argument deserves a straight answer. Nobody keeps a card on file for a tool they ignore, and paying is a harder signal than answering a survey. True, and it’s why we read the index as real evidence of where buying attention has gone. A payment still can’t say whether the tool was used on a Tuesday or sat idle by June.
Is the price per task rising while the lead grows?
Possibly, and Ramp says so itself. Its write-up flags frequent outages, rate limits and growing dissatisfaction with results, and says a recent Anthropic model update would triple token costs for any prompt that includes an image. A vendor gaining customers while the cost of a task climbs is a pricing story as much as a popularity one.
Sonnet 4.6’s launch pricing showed the same gap between headline and task cost. Ramp also names an incentive problem. Anthropic earns more as businesses buy more tokens, so it may steer buyers toward pricier models when a cheaper one would do, and Ramp says OpenAI feels the same pull. That’s an observation from the index’s author, and also just how usage billing works.
On May 28, 2026, Anthropic said Opus 4.8 would keep the list price of Opus 4.7, $5 per million input tokens and $25 per million output tokens. A steady list price is welcome. It doesn’t fix a prompt that suddenly needs three times the tokens, which is what our look at GPT-5.5 pricing found on the OpenAI side too.
The second-order effect matters more than the first. If the leader’s customers pay more per task and hit more rate limits, the lead may be partly a queue of buyers about to start comparing. Whether April was a swing or a trend will show in Ramp’s next monthly index. Watch that and your own invoices, because the Microsoft and OpenAI arrangement already hints that seat and metered pricing are in motion.
How should you read your own AI bill?
Pull 90 days of AI charges from the company card and the accounting system, then tag each by vendor and by team. Split flat seats from metered usage, because seats are predictable and metered usage is where bills surprise you. Then run ten real tasks through your top two vendors and record time and cost.
Then take your top two workflows and run the same ten real tasks through each vendor. Record the minutes to an acceptable result and the cost. That’s an afternoon of work, and it produces the number the index can’t.
Here’s a hypothetical, ours, to show the arithmetic. A firm finds 40 seats on one assistant and six on another, plus a metered coding tool that cost $900 last month. If the metered tool does the work of two junior hires on one project, that line item deserves the scrutiny, not the 40 seats. A leaderboard would count three vendors and tell you nothing about it. For a wider frame on totals, see our guide to what AI costs a business, and for a side-by-side of the main assistants, our ChatGPT, Claude and Gemini comparison.
Read the results in two columns, cost per accepted task and minutes to an acceptable result. Say vendor B is 20% cheaper per token but needs two extra rounds of correction on every task. The cheaper line on the invoice is the dearer one in staff time. That second-order cost is invisible to an index of payments, and it may well be larger than the seat price. Count your people’s minutes as well as the vendor’s meter. Hardware costs reach the invoice later, as the Nvidia price increase shows.
Ramp hasn’t published its underlying data, so if it releases usage depth or a sample that mirrors the wider market, we’d give the lead more weight. One more wrinkle: the index counts paid accounts, so a company that stops using a tool but forgets to cancel still shows up. Pick vendors from your invoices and a ten-task test. Book the afternoon this week.
Frequently asked questions
What is the Ramp AI Index?
It’s a monthly measure Ramp builds from corporate card and invoice payments on its own platform. It shows the share of businesses paying for each AI vendor, plus overall paid AI adoption, which was 50.6% in April.
Did Anthropic overtake OpenAI in business AI adoption?
On Ramp’s spend measure, yes. Ramp put Anthropic at 34.4% of businesses and OpenAI at 32.3% in April. That counts who pays, not who has the most users or does the most work, so it isn’t a full market-share ranking.
Should a small business switch AI vendors because of the Ramp AI Index?
No, not on that evidence alone. Total your own AI invoices for 90 days, split seats from metered usage, and test your two biggest workflows on both vendors before you change anything.
Archive entry dated May 14, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
