Source: Ramp Economics Lab — 2026-08-12
Summary
Ramp's August 2026 AI Index, built from anonymized corporate card and bill-pay data across more than 70,000 U.S. businesses, shows Anthropic leading business AI adoption at 43.5% of businesses (up 1.1 points month-over-month) versus OpenAI at 39.7% (up only 0.23 points). The report's headline argument is that frontier pricing is hitting a ceiling: Anthropic's flagship Fable 5 model, despite being by far its most expensive offering, accounts for only 6% of Anthropic tokens purchased but 11.4% of dollars spent, suggesting businesses are increasingly unwilling to pay a premium for marginal capability gains.
Key Takeaways
- 43.5% vs. 39.7% is real, spend-based adoption data — not survey self-reporting — making Ramp's index one of the more grounded signals available on which AI vendor businesses are actually paying for, at scale, across 70,000+ companies.
- The month-over-month growth gap (Anthropic +1.1 points vs. OpenAI +0.23 points) suggests Anthropic isn't just ahead, it's still pulling further ahead, while OpenAI's overall business-adoption growth has slowed noticeably.
- The Fable 5 pricing data point — 6% of tokens but 11.4% of dollars — is the clearest evidence in the report that a real price ceiling is forming: businesses are increasingly routing routine work to cheaper models and reserving the most expensive frontier model for a narrow slice of tasks where it's worth the premium.
- xAI posted its fastest growth since July 2025 (up 0.94 points to 4% of businesses), and open-source/model-serving platforms rose to 6.1% of businesses (up 0.2 points) — both signs that budget-conscious buyers are increasingly willing to look past the two market leaders.
Reel Script
Hook (17s, ~38 words): Real corporate spending data from over 70,000 businesses just showed Anthropic pulling ahead of OpenAI in the enterprise AI race — and buried in the same report is evidence that businesses are starting to say no to the most expensive models.
Core Concept (75s, ~170 words): Most "who's winning the AI race" narratives are built on hype, funding announcements, or self-reported surveys. Ramp's index is different — it's derived from actual corporate card and bill-pay transactions across more than 70,000 U.S. businesses, so it measures what companies are literally paying for, not what they say they're excited about. On that measure, Anthropic now leads at 43.5% of businesses paying for its subscriptions or token usage, versus 39.7% for OpenAI. What makes this more than a snapshot is the momentum: Anthropic grew 1.1 percentage points in a single month, while OpenAI grew only 0.23 points — meaning the gap isn't static, it's widening. That's a meaningful shift from a year where OpenAI was the default assumption for enterprise AI spend, and it's happening through real dollars changing hands, not press coverage.
Hands-On (70s, ~160 words): The most interesting data point isn't the vendor race — it's what's happening within Anthropic's own product line. Fable 5, Anthropic's flagship and by far its most expensive model, makes up only 6% of the total tokens businesses purchase from Anthropic. But it accounts for 11.4% of the dollars they spend. Do the math on that gap: businesses are paying disproportionately more per token for Fable 5 while using it sparingly, and routing the bulk of their actual token volume to cheaper models in Anthropic's lineup instead. That's a price-ceiling signal in the data itself — businesses aren't rejecting the top-tier model outright, they're rationing it to the specific tasks where its extra capability is worth the extra cost, and defaulting to cheaper models everywhere else. Add in xAI's fastest growth since mid-2025 and open-source/model-serving platforms climbing to 6.1% of businesses, and you get a market that's diversifying downward on price, not just consolidating around the two leaders.
Takeaway (22s, ~48 words): If you're budgeting AI spend, the market itself is already doing this: don't default every task to the most expensive model. Ramp's data says most businesses have quietly figured out which tasks actually need frontier capability — and which don't.