Two months after Anthropic shipped a model so capable its release was briefly blocked by the U.S. government, American businesses have returned a verdict that has almost nothing to do with capability: they will not pay for it.
Fable 5 accounts for roughly 11% of what companies spend on Anthropic's tools, according to spend data from Ramp, the payments company that tracks purchasing across more than 70,000 businesses. In July, the flagship supplied just 6% of the Anthropic tokens Ramp could identify while absorbing 11.4% of Anthropic-attributed dollars. By August 23 — more than two months after launch — that dollar share had settled near 11% and stopped climbing. Ramp told the Financial Times that Claude Opus 5, released July 24 at half the price, had already overtaken it.
The external comparison is the one that stings. OpenAI's flagship, GPT-5.6 Sol, takes 25% of OpenAI tokens and 23% of OpenAI spend in the same dataset. In July, Fable 5 generated only about 75% as much model-attributed spending as Sol — the best model on the market losing, on the P&L, to the second-best.
“Most people don't need to operate at the frontier,” Miles Clements, a partner at Accel, which has put close to $1 billion into Anthropic, told the Financial Times. The stretch when corporate buyers reflexively reached for the most advanced option available, he said, “was not a durable era.”
The arithmetic of an acceptable answer
Price explains most of the gap. Fable 5 lists at $10 per million input tokens and $50 per million output — roughly twice GPT-5.6 Sol and, more damagingly, twice Anthropic's own Opus 5, which launched July 24 at $5 and $25. Anthropic's evaluations show Opus matching or beating Fable on bounded coding and office work at a lower cost per completed task. Dianne Penn, Anthropic's product leader, has effectively conceded the segmentation, telling customers to choose Opus 5 for value and reserve Fable 5 for “days-long, very autonomous projects.”
That is a reasonable product story. It is also an admission that the frontier is now a niche. Enterprises have built routing layers that dispatch routine work to whatever is cheapest and escalate only what fails — a dispatch rule, not a vendor commitment. Among Vercel gateway teams running more than 10 million tokens in both June and July, three in four changed at least a tenth of their model mix in a single month. Switching a model there takes one line of code.
OpenAI is pressing exactly that seam. On August 21 it cut GPT-5.6 Sol's developer pricing by more than 20%, to $4 input and $20 output per million tokens from $5 and $30, running through at least November 21. Mantas Lukauskas, AI tech lead at the website host Hostinger, described the round of cuts as the “first real test” of whether U.S. labs can defend pricing on their most advanced products.
The bearish read is not unanimous. Vercel's July gateway index put Fable at 13.2% of all spending, second only to Opus 4.8, with nine in ten Fable teams new to the product — and Anthropic collecting 65.1% of gateway spend on 30% of tokens, at 4.4 times the average price per token across other labs. Ramp's sample skews toward technology firms; Vercel values traffic at list prices rather than negotiated bills. Neither settles the question — but their disagreement establishes that the mix moves fast.
What commoditization at the frontier does to lab economics
The frontier premium has always been the load-bearing assumption in AI lab finance: spend billions on the next training run, recover it by charging a multiple for the resulting capability edge. Ramp's data is the first clean evidence that the multiple has a ceiling.
“With Fable 5, we've found a new upper bound for how much businesses are willing to spend on AI,” wrote Ara Kharazian, Ramp's lead economist, in the August index. “Here, more performance is not worth the price tag.” To move buyers up the stack, he argued, labs will now have to clear two bars at once — deliver capability beyond even Fable 5, and keep competitors from getting reasonably close. With open-weight systems trailing by only a few months, that is “increasingly out of reach.”
The structural pressure is visible underneath. Open-weight models ran 36% of Vercel gateway token volume on 8.6% of spend in July, up from 29% of volume on under 4% in June. Average price per token across the gateway fell 13.6% in a month. On Ramp's broader sample, 6.1% of AI-using businesses now buy through model-serving platforms that front open-source and Chinese models — small, but rising monthly and concentrated among the heaviest spenders.
None of this has dented Anthropic's top line yet. Its annualized revenue rate reached $65 billion in July, up from $47 billion in May, on preliminary second-quarter revenue above $11.5 billion and positive adjusted operating income. Ramp found 43.5% of U.S. businesses paying Anthropic in July against 39.7% for OpenAI. The company is winning the provider war while losing the mix war inside its own catalog — growth that increasingly comes from cheaper SKUs sold to more customers, not premium SKUs sold to the same ones.
That is a materially different business: it compounds more slowly and makes every incremental training run harder to underwrite.
What to watch
Three things through the fall. First, whether Opus 5's share keeps climbing at Fable's expense once Ramp publishes September model-level data — a durable inversion inside Anthropic's own lineup would confirm the ceiling rather than a slow start. Second, whether OpenAI's three-month Sol discount becomes permanent on November 21; a quiet extension would signal that list prices at the frontier are now fiction. Third, the top-of-market spenders: Ramp's top 1% of businesses spent a median $7,400 per employee on AI in July, against $11.95 for the median firm. If that cohort's spending flattens while its token volume keeps rising, the commoditization thesis stops being a pricing story and becomes a revenue one.
“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.”— Ara Kharazian, Lead Economist, Ramp