Claude Fable 5, the highest-scoring language model on the market, made up only 6 percent of the tokens businesses bought from Anthropic in July and 11.4 percent of the dollars they spent there, according to spend data published by the corporate card company Ramp. Anthropic's overall business adoption rose over the same period. The story is not that customers are leaving Anthropic -- it is that they are refusing to pay for its best model.
Key facts
- Fable 5 accounted for 6 percent of Anthropic tokens purchased and 11.4 percent of Anthropic model spend in July 2026.
- OpenAI's flagship, GPT-5.6 Sol, took 25 percent of OpenAI tokens and 23 percent of spend over the same window.
- Published August 12, 2026 by Ara Kharazian, lead economist at Ramp, using anonymized spend data from U.S. businesses.
- Primary source: Ramp's August 2026 AI Index.
For most of the last three years, the assumption underneath the AI industry's economics was that a better model commands a higher price, and buyers will pay it. Ramp's July numbers are the first clear public evidence that the assumption has a ceiling. Kharazian puts it bluntly: "So with Fable 5, we've found a new upper bound for how much businesses are willing to spend on AI. Here, more performance is not worth the price tag."
The comparison inside Anthropic's own catalog is what makes the point. Fable 5 lists at roughly $10 per million input tokens, about twice GPT-5.6 Sol. Ramp finds that despite being the most expensive model by a wide margin, Fable 5 generated only about 75 percent as much model-attributed spend in July as OpenAI's flagship did -- so it is not simply that a premium model captures a small share of a large base. It is capturing less money in absolute terms than a cheaper competitor.
The mechanism is straightforward once you look at cost per unit of work rather than cost per token. Independent benchmarking by Artificial Analysis scores Fable 5 at 62 on its intelligence index at $3.14 per benchmark task, while Anthropic's own Opus 5 scores 63 at $2.34. The best model on the leaderboard is not the best buy on the invoice -- and it is not even the best buy within its own vendor's lineup. Public routing traffic tells the same story: OpenRouter's Anthropic page shows roughly 1.47 trillion tokens processed through Opus 5 against about 272 billion through Fable 5, with the usual caveat that OpenRouter measures its own pipe, not the market.
There is a genuine confounder, and it deserves stating. Fable 5's first weeks were disrupted: Anthropic launched it on June 9, suspended access on June 12 under a U.S. government export directive, and restored it on July 1, with subscription plans capped at half of weekly usage limits through July 7 before moving to usage credits. A model that was unavailable for most of a month will look under-adopted in the month that follows. Ramp's data covers the period right after that gap.
Why it matters: the market Ramp describes is one where the marginal frontier model has to prove not just that it is better, but that it is better by more than the price difference -- and Kharazian argues that bar is getting harder to clear "especially as open source models catch up to being only a few months behind." Ramp's own numbers show 6.1 percent of AI-spending businesses now using model-serving and inference platforms, the on-ramp for open-weight models, up from 4.5 percent in January. That is the substitution pressure, arriving slowly and from underneath. It is also the backdrop against which both frontier labs have filed to go public.
The caveat: spend share is not the same as usefulness, and a small number of very expensive queries can matter enormously to the buyer running them. Ramp's sample is also self-selected -- companies that use Ramp -- and skews technical, which Kharazian says means true Fable 5 adoption is probably lower than his estimate rather than higher. And Anthropic itself is not in trouble by any public measure: it reported run-rate revenue crossing $47 billion in its Series H announcement in May. What is under pressure is a pricing theory, not a company.
Originally published on Ground Truth, where every claim is checked against the primary source.
