What does Ramp’s data on more than 70,000 U.S. businesses reveal about the changing competition between OpenAI and Anthropic among paying corporate users—including their market shares from May through July, OpenAI’s faster Q3 to date growth, the data’s limitations and tech sector

Ramp’s data suggests that Anthropic still led OpenAI among its U.S. business customers in July, but OpenAI had begun growing faster in Q3 to date—evidence of a fluid contest rather than a settled enterprise-AI market. The findings are a useful spending signal, not a national market-share measure. 67

What the shares show

• In May, Anthropic overtook OpenAI: about 41% of Ramp businesses paid Anthropic, versus about 39% paying OpenAI. 5
• By July, Anthropic’s lead had widened to 43.5% of businesses, while OpenAI reached 39.7%. 3
• Yet more-current Q3-to-date Ramp data indicated that OpenAI was adding paying business users faster than Anthropic, beginning to narrow that gap. 67

These are adoption shares—the proportion of Ramp customers paying each provider—not mutually exclusive revenue shares. A company can pay both vendors.

Why the apparent reversal may be happening

• Ramp’s July model-usage data shows OpenAI’s GPT-5.6 Sol accounted for 25% of its tokens and 23% of its model spending, versus Fable 5’s 6% of Anthropic tokens and 11.4% of Anthropic model spend. Fable generated roughly 75% as much attributed spending as Sol. 3
• One plausible explanation is developer preference for Sol’s price–performance and usability. Ramp found Fable cost about $10 per million tokens—roughly twice Sol’s price—despite its stronger reported performance. 3
• A further potential obstacle for regulated or privacy-sensitive customers was Anthropic’s 30-day retention requirement for its most advanced models. Reporting said Anthropic planned to give enterprise customers greater control over that policy, so this was a material but potentially temporary disadvantage. 124

These are plausible explanations, not causal proof: Ramp’s transaction data cannot establish why a firm selected, added, or discontinued a model.

Important limitations

• The sample covers aggregated, anonymized card and bill-pay transactions from more than 70,000 U.S. Ramp businesses, rather than all U.S. companies or all enterprise AI contracts. 3
• It likely underrepresents spending paid through cloud marketplaces, resellers, procurement systems, bundled software agreements, or invoices outside Ramp.
• The token-spend sample used for the Fable-versus-Sol comparison is somewhat more technology-oriented than Ramp’s ordinary AI Index sample; Ramp says actual Fable adoption may therefore be lower than its estimate. 3
• More broadly, Ramp customers and the firms purchasing direct model tokens are likely more tech-forward than the typical U.S. business. The results should not be described as definitive national enterprise market share. 43

What the volatility means

• Anthropic’s fast rise followed by OpenAI’s faster Q3-to-date growth implies that enterprise AI expenditure is not yet especially sticky at the vendor level. Firms appear willing to trial, multi-source, or switch models as performance, price, privacy terms, and developer tooling change. 67
• That does not mean AI spending itself is fading. It means vendor lock-in remains limited and product releases can still quickly shift purchasing behavior.

The market itself is expanding

• Paid AI adoption on Ramp crossed 50% in March, reaching 50.4% of businesses. 3
• By July, the share was nearly 56%, meaning the competitive battle is occurring inside a rapidly enlarging buyer base, rather than being merely a transfer of a fixed set of corporate customers. 3

The central takeaway is therefore two-sided: Anthropic retained the larger July footprint in Ramp’s sample, while OpenAI’s recent momentum hints that pricing, privacy, and developer experience can still rapidly reshape corporate AI purchasing.