Businesses Just Told Anthropic No Thanks on Fable 5, and Nobody Had to Legislate It

Ramp tracks how more than 70,000 US businesses actually pay for AI, through the corporate cards it issues them, and its August AI Index just published a number that should worry Anthropic more than any competitor's press release. One month after launching Fable 5, the model widely regarded as the most capable on the market, it accounted for only 6 percent of Anthropic's token volume and 11.4 percent of dollars spent on Anthropic models among Ramp's customers. OpenAI's GPT-5.6 Sol, priced at roughly $5 per million tokens against Fable 5's roughly $10, pulled 25 percent of OpenAI's tokens and 23 percent of its spend over the same kind of window.

That is not a small gap. It is a business's best model, by capability, losing the adoption race to a rival's mid-tier option priced at half the rate. Ramp economist Ara Kharazian's read is direct: businesses are showing an upper bound on what they will pay for marginal performance gains. Anthropic still leads the category overall, at 43.5 percent of Ramp-tracked businesses paying for its products in July versus OpenAI's 39.7 percent, both up slightly month over month. But the model-level split says the premium tier is where the resistance is, not the vendor relationship itself.

Here is the steelman for reading this differently. Two vendors holding a combined 83 percent of business AI spend is not a wide-open market with dozens of live options disciplining every price point. It looks more like a duopoly where one giant is undercutting another giant on price, and calling that "the free market at work" risks giving competition more credit than two companies checking each other's pricing deserves. TechCrunch's coverage of the same data adds that OpenAI is now growing faster than Anthropic among business customers in the current quarter, which could just as easily mean buyers are consolidating around whichever giant is cheaper this month, not exercising real choice across a broad field. And Ramp's own economist flags a real methodology caveat worth taking seriously: the sample behind the token-level Fable 5 numbers skews toward more technical companies than Ramp's customer base as a whole, so if anything the real adoption gap among ordinary small businesses is probably wider, not narrower, than 6 percent.

Both of those points are fair, and I would rather state them than pretend this is a tidier story than it is. But neither one changes what actually happened here. Nobody capped Fable 5's price. No agency required Anthropic to publish a value comparison against GPT-5.6 Sol. Tens of thousands of businesses looked at what the top-tier model cost against what a cheaper model delivered for their actual workload, and enough of them said the extra dollars were not worth it that the gap shows up clearly in Ramp's numbers a month after launch. That is buyers pricing quality against cost with their own money, which is exactly the mechanism a free market runs on, duopoly or not. A genuine monopolist does not need to worry about this kind of signal because customers have nowhere else to route their spend. Here they had somewhere else, and 19 percentage points of token share went there instead. Worth noting too: this is not strictly a two-horse race. xAI picked up 0.94 points of business share in the same period to reach 4 percent, and open-source and model-serving platforms grew to 6.1 percent, both real alternative destinations for the spend that did not go to Fable 5.

The other detail in Ramp's index matters more than the headline. The share of Ramp-tracked businesses paying for any AI product climbed to nearly 56 percent in July, up from just over 50 percent in March. Total demand for AI is still growing. What is not growing at the same clip is tolerance for paying double the price for a model that beats its cheaper rival on benchmarks but not by enough to justify the invoice on routine work. That is the distinction that should matter to any business owner reading this, not "which vendor is winning" but "which model tier is actually worth the premium for what I run through it."

This is the same discipline I have written about with model routing: most of what a business sends through an AI system is routine enough that the frontier model's extra reasoning ability never gets used, so paying frontier prices for it is pure waste. Ramp's data says the market has started doing that math in aggregate, at scale, without anyone telling it to. If you are still defaulting every request to your most expensive model out of habit rather than a measured need, you are the reason Fable 5's adoption curve looks the way it does, and you are leaving money on the table the same way. Tell me what you are running and I will help you figure out where the premium model actually earns its price and where it does not.