Hold Ramp's two statements against each other and they scrape. In May the standings were 41 to 39, a two-point Anthropic lead [2][1]; by July they were nearly 44 to nearly 40, about four points [3][2], so on the published prints the gap doubled rather than closed [3]. Ara Kharazian, Ramp's economist, nonetheless says OpenAI is the faster grower in Q3 to date [5]. July is itself a Q3 month, which means the closing has to be happening in the three weeks of August that Ramp has not printed, on a quarter with a month left to run [6]. That is a real signal, or rather it is a real signal about August specifically, and it is not the same claim as the four-point lead everyone will quote.
The more interesting number is the one neither lab is fighting over. Between May and July the two of them went from 80 per cent of Ramp's paying AI customers to about 84 [4][5], four points off everyone else, of which Anthropic took roughly three and OpenAI one [6]. Anthropic's lead, expressed against the only competitor that matters here, is 52 per cent of a two-horse pool [8]. That is a coin flip with a lean.
Underneath the duel, the pool grew. Companies paying for any AI in Ramp's base topped 50 per cent in March and hit nearly 56 by July [12], which is about 12 per cent more paying customers in four months [7]. A lab can drop a point of share into that and still add logos, so the share series and the revenue series can point opposite ways for a couple of quarters without either being wrong.
What the data cannot do is tell you what a point is worth. Ramp gave percentages and withheld the dollars [7], the sample skews toward tech [4], and it excludes large enterprises that run spend management through providers such as American Express [8]. A card-visible share of buyers is a price of attention; the value sits in commitments that a corporate card never sees. TechCrunch reads the flip-flopping as a warning to both companies' investors about how sticky enterprise AI spending really is [13], and that is the right worry pointed at possibly the wrong layer.
This is probably wrong, but my read is that the volatility is concentrated where switching costs are a configuration change, and Kharazian's own explanation supports it: he credits GPT-5.6 Sol with becoming the developer default and blames Fable 5's price plus regulator-imposed data retention for weak adoption [9], which is how a self-serve buyer behaves rather than a procurement committee. Anthropic's top tier is expensive by design and aimed at narrower use cases [10], and the 30-day retention notice it gave Fable users drew open anger [11]. The counter-thesis is straightforward and might well win: if a dollar-denominated series ever shows the same four-point ordering, then Ramp's count of buyers was a fine proxy all along and my layer objection dies. We only get that series when these two get close enough to their planned IPOs to open the books [14].
