The AI trade's weak link is the buyer: Anthropic's best model took 6% of its tokens

The arithmetic Ramp's letter leaves on the table is the useful part. Fable 5 sells for roughly $10 per million tokens, twice the price of GPT-5.6 Sol, which Ramp still describes as highly performant ^[9]. Over July, Fable generated about 75% as much model-attributed spend as Sol ^[8]. Divide the spend ratio by the price ratio and Fable's token volume is running at roughly 37% of Sol's ^[15]. The most performant model on the market, by Ramp's own account ^[19], is being metered by its buyers rather than turned down.

Chain the same figures the other way and Anthropic looks stronger than its adoption lead implies. Fable's share of Anthropic dollars, Sol's share of OpenAI dollars, and the ratio between the two imply that total model-attributed spend at Anthropic is running about 1.5 times OpenAI's inside Ramp's token-level panel ^[20]. That is a wider gap than the 3.8 point lead in business count ^[13], which itself widened by 0.87 points in a single month ^[14]. Ramp flags that this panel skews more tech-heavy than its usual sample ^[10], a caveat that trims Fable's adoption figure and Anthropic's dollar position at the same time.

The mechanism behind the ceiling is worth stating in Ramp's own terms, because it is narrower than "AI demand is slowing." First-time AI buyers are still going to the American labs, not to open source or Chinese models ^[5]. Because headline adoption growth at OpenAI and Anthropic has slowed, more of their future revenue has to come from businesses that already buy, and it is the advanced spenders in that group who are moving onto model serving platforms ^[18], now used by 6.1% of AI-using businesses and still climbing ^[4]. The accounts with the largest budgets are the accounts with the most substitution capacity. Fable 5, priced at double a good-enough rival, is the clean test of what those budgets will bear, and Ramp reads the result as a new upper bound on willingness to pay ^[17].

Underneath the spend totals sits a distribution that decides how much any of this matters: a median of $7,400 per employee at the top 1% of firms in July, $650 at the top 10%, and $11.95 at the median firm ^[11]. The top percentile is therefore spending roughly 620 times the median firm per head ^[16]. Growth in aggregate business AI spend is a story about procurement at a very small cohort of companies, and that cohort is the same one buying open weight capacity by the token. Model quality is not the variable under pressure here. The line item is.