43.5% of US businesses pay Anthropic. The median one spends $11.95 per employee.

TL;DR: Ramp published its August 2026 AI Index this week, and the headline everyone ran is the scoreboard: Anthropic at 43.5% of US businesses (up 1.1 points month-over-month), OpenAI at 39.7% (up just 0.23), xAI at 4%, and open-weight serving platforms at 6.1% of AI-using firms. Anthropic leads; OpenAI is growing faster in Q3 to date. That is a real finding and a small one. The number that matters sits further down the release: the median AI-buying firm spends $11.95 per employee, against $650 for the top 10% and $7,400 for the top 1%. That is roughly a 620x spread between firms that have deployed AI and firms that have expensed it. Ramp’s own framing — lead economist Ara Kharazian titled the release “Cracks in the AI thesis” — focuses on Fable 5 taking only 6% of Anthropic tokens despite being the premium tier, and reads it as a ceiling on what buyers will pay for capability. That reading is half right. The other half is that Anthropic undercut Fable 5 itself on 24 July with Opus 5 at half the price. And one more thing worth holding: Ramp published this a day after launching a free model router whose entire pitch is that switching vendors should be cheap.

Breadth is nearly saturated. Depth has barely started.

The two percentages that led every writeup of this index are breadth measures. They count the share of businesses in Ramp’s panel with any line item from a given vendor. A 300-person firm with four Claude Pro seats and a 300-person firm running Claude across every engineering workflow are one unit each.

Once you know that, the movement in those numbers reads differently. Anthropic added 1.1 points month-over-month. OpenAI added 0.23. In a market that is supposedly in the steepest part of its adoption curve, the two dominant vendors moved a combined 1.3 points in a month. That is not a land grab. That is a market where nearly everyone who is going to try AI has already tried it, and the remaining question is how much they do with it.

Which brings us to the depth number:

Segment	Median AI spend per employee
Top 1% of firms	$7,400
Top 10% of firms	$650
Median AI-buying firm	$11.95

Twelve dollars per employee per year — not per month. At list price that is roughly one Claude Pro or ChatGPT Plus seat for every fifteen employees, or a trivial API bill. The typical company in this dataset has adopted AI in the sense that it has a receipt.

The 620x gap between the median and the top percentile is the actual story of enterprise AI in 2026. There is a thin band of genuinely AI-native companies spending like the technology is load-bearing, and a very long tail that has bought a few seats and stopped. Every projection built on “X% of businesses have adopted AI” is measuring the tail and pricing the head.

The Fable 5 result is being misread

Kharazian’s central exhibit for the “cracks” thesis is Fable 5, Anthropic’s premium frontier tier. In the July data it took 6% of tokens businesses bought from Anthropic and 11.4% of dollars — real money, about 75% as much spend as GPT-5.6 Sol generated for OpenAI, but a small slice of a big vendor. At roughly $10 per million input tokens, double GPT-5.6 Sol’s $5, the natural read is that buyers looked at the premium and declined.

Two facts complicate that.

Fable 5 barely had a July. US export controls pulled Fable 5 and Mythos 5 offline for 18 days in June. Anthropic restored the model globally on 1 July, initially bundled into subscription weekly limits rather than sold straight. The July window is close to Fable 5’s first clean month of unrestricted availability, measured against models that had been shipping continuously for a quarter.

Anthropic undercut it mid-month. On 24 July, Anthropic launched Claude Opus 5 at $5/$25 per million tokens — half Fable 5’s price, matching GPT-5.6 Sol on input and beating it on output, and outperforming Fable 5 on several benchmarks. Buyers evaluating Anthropic’s premium tier in late July were not choosing between Fable 5 and nothing. They were choosing between Fable 5 and a cheaper, arguably better model from the same vendor, released while they were deciding.

So the ceiling is real, but it is narrower than “businesses won’t pay for capability.” It is: buyers will not pay a premium for capability that gets repriced within weeks — and on this occasion the vendor repriced it before the buyers finished evaluating. Anthropic conceded the point faster than the market could make it. That pattern is now general across the industry, as the August price war coverage documented: frontier pricing is not a moat, it is a temporary position.

The commoditisation is coming from below

The quieter figure in the index is 6.1% of AI-using businesses now paying a model-serving platform for open-weight models, up 0.2 points in a month. That is a small number attached to a specific kind of buyer — the ones sophisticated enough to run Kimi K3 or GLM-5.3 on rented hardware rather than buying tokens from a lab.

That segment overlaps heavily with the high-spend band. The firms spending $7,400 per employee are the firms with the engineering capacity to arbitrage inference, and they are the firms whose spend the frontier labs most want. Commoditisation rarely arrives as a collapse in the headline price. It arrives as the most valuable customers quietly moving the highest-volume workloads somewhere cheaper and keeping the frontier tier for the hard 10%.

The scorekeeper is now a player

One disclosure belongs in any reading of this index. On 19 August, Ramp launched Router, a model gateway that is free through the end of 2026 and whose stated premise is that the best model changes constantly and switching should be costless. On 20 August, Ramp published an index arguing that businesses “flop back and forth as each lab releases new models” and that investors should question how sticky enterprise AI spending really is.

This is not an accusation. The underlying data is Ramp’s own transaction ledger, the methodology is disclosed, and the findings are unhelpful to plenty of parties Ramp has no reason to target — including the general AI-spend growth thesis that Ramp’s own product line benefits from. Kharazian has been publishing this index since well before Router existed.

But emphasis is a choice, and a vendor’s market research will foreground the market structure its product addresses. The volatility narrative is the strongest available argument for buying a router. Read the numbers as evidence and the narrative as a hypothesis, the same way you would read a security vendor’s threat report.

What to actually do with this

Stop quoting adoption percentages as if they mean deployment. If you are building a business case internally, the $11.95 median is the honest benchmark for where most companies are, and the $650 top-decile figure is the honest benchmark for what “we are serious about this” costs. Anyone citing “43.5% of businesses use Anthropic” as evidence of market maturity is citing a receipt count.

Do not standardise on a vendor. A four-point lead narrowing while a third vendor grows a point a month off a 4% base is not a stable market. Keep model calls behind an OpenAI-compatible interface so the vendor is a config value — the same architectural advice that the gateway consolidation made urgent, now with demand-side data behind it. Compare Claude and ChatGPT per workload, not per company.

Keep your own spend telemetry. Ramp can see this because it holds the transactions. Inside your own company, the equivalent visibility is token-level logging you own, not a dashboard your gateway or your lab provides. It is what lets you price a migration without asking the incumbent for the numbers.

Treat premium tiers as rentals. The Fable 5 result is the cleanest evidence yet that the top of a vendor’s lineup is a temporary price point. Do not architect around a frontier model’s specific capabilities unless the workload genuinely requires them, and assume the price you are paying today is the highest you will ever pay for that tier — including from the vendor who sold it to you.

For the current state of the tooling this data describes, see the best AI coding tools roundup and the individual reviews for Claude, ChatGPT and Grok.

Sources

Frequently asked questions

Is the Ramp AI Index a reliable measure of enterprise AI adoption?

It is reliable for exactly what it measures and routinely over-read beyond that. Ramp tracks real transactions — corporate card charges, bill pay and token spend — across more than 70,000 US businesses, which makes it far better evidence than a survey of stated intent. Nobody misreports a credit card charge. The limits are equally concrete. Ramp says its panel skews tech-heavy, so it runs ahead of the broader economy. It excludes large enterprises that pay via American Express, invoicing or negotiated committed-spend contracts, which is precisely how the biggest AI deals are actually papered. And it counts a business as having adopted a vendor if that vendor appears on the bill at all, so a single $20 seat and a seven-figure commitment are one unit each. Treat it as the best available high-frequency signal on the small and mid-market, not as a census of enterprise AI.

If 43.5% of businesses pay Anthropic, why is the median spend only $11.95 per employee?

Because those two figures answer different questions, and only the second one is about deployment. The 43.5% is a breadth measure: the share of firms with any Anthropic line item. The $11.95 is a depth measure: what the typical adopting firm actually spends, normalised per head. A 200-person company running two $20 Claude seats clears the breadth bar and lands near the depth floor. The distribution is what makes this stark — the top 1% of firms spend a median of $7,400 per employee and the top 10% spend $650, so the mean is dragged upward by a thin band of genuinely AI-native companies while the median firm has barely started. The honest summary is that adoption is nearly universal and usage is nearly nonexistent, and most commentary about the AI market conflates the two.

Does Fable 5's weak uptake prove businesses have hit a price ceiling?

Partly, but the cleaner explanation is that Anthropic undercut its own flagship. Fable 5 took 6% of Anthropic tokens and 11.4% of Anthropic dollars in the July data, which does look like buyers refusing to pay roughly $10 per million input tokens. Two things complicate the reading. Fable 5 spent much of the preceding period unavailable — US export controls pulled it offline for 18 days in June, and it was only restored globally on 1 July, so the July window is close to its first clean month of general availability. Then on 24 July Anthropic shipped Claude Opus 5 at $5/$25 per million tokens, half Fable 5's price, beating it on several benchmarks. Buyers were not only declining a premium tier; they were being offered a cheaper one by the same vendor mid-quarter. The ceiling is real, but it is a ceiling on paying a premium for capability that gets repriced within weeks — and Anthropic agreed with the market before the market finished voting.

Should the fact that Ramp sells a model router change how I read its index?

It should change your reading of the framing, not the numbers. The transaction data is Ramp's own ledger and there is no obvious way to skew it; the methodology is disclosed and the findings are unflattering to plenty of parties, including the general AI-spend thesis Ramp benefits from. What is worth noticing is timing and emphasis. Ramp launched Router — a free-through-2026 gateway whose entire pitch is that switching models should be cheap — on 19 August, and published an index arguing that businesses flop between labs and that enterprise spending is not sticky a day later. Both things can be true simultaneously and probably are. But a vendor's market-structure research will tend to foreground the structure its product addresses, so take the volatility narrative as a hypothesis supported by good data rather than as a neutral reading of it.

Anthropic leads on adoption but OpenAI is growing faster. Which one should I standardise on?

Neither, if standardising means writing either vendor's SDK into your application layer. That is the actual lesson of a dataset showing a four-point lead narrowing while both vendors add share and a third — xAI, at 4% and up nearly a point in a month — grows off a small base. The leaderboard is changing faster than a procurement cycle completes, so any architecture that treats a vendor choice as permanent will be wrong before it is finished. Keep model calls behind an OpenAI-compatible interface so the vendor is a configuration value, keep your own token-level spend logs so you can price a migration without asking the incumbent, and pick per workload rather than per company. Standardise the interface, not the supplier.

What does 'cracks in the AI thesis' actually refer to?

Ramp economist Ara Kharazian is pointing at the gap between the industry's revenue assumptions and the observed willingness to pay. The bull case for frontier labs requires that each capability jump commands a price premium and that spending compounds as models improve. The July data undercuts both halves. The most expensive frontier tier took a small minority of tokens, growth in the share of businesses adopting AI is slowing to fractions of a point per month at both leading vendors, and the more sophisticated buyers are increasingly routing work to open-weight models served on third-party platforms — 6.1% of AI-using businesses now pay one. That is the shape of a market commoditising from the bottom while the top tier struggles to hold a premium. It is not evidence that AI spending is falling. It is evidence that it may not compound the way the valuations assume.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.