Fable 5 Took Just 6 Percent of Anthropic's Tokens

What happens when the best model on the market goes on sale and corporate buyers mostly shrug?

That is roughly the story in Ramp's August 2026 data. Fable 5, Anthropic's most capable model, accounted for just 6 percent of the tokens companies purchased from Anthropic during its first month on sale. Measured by money rather than volume it does better, at 11.4 percent of total spending on Anthropic models, which is another way of saying it is expensive rather than popular.

The comparison that stings

Despite costing considerably more per token, Fable 5 generated roughly 75 percent of the revenue that OpenAI's GPT-5.6 Sol produced. A premium price failed to compensate for thin volume.

The pricing explains a lot. Fable 5 runs $10 per million input tokens and $50 per million output, about twice GPT-5.6 Sol and about twice Anthropic's other flagship models.

None of this reflects a struggling vendor. In July 2026, 43.5 percent of US companies paid Anthropic for subscriptions or tokens, against 39.7 percent for OpenAI, whose month-over-month growth was a nearly flat 0.23 points. Anthropic is not having trouble acquiring customers. It is having trouble selling those customers the top of its range.

A gap of several hundred times

One more figure from the same dataset reframes the whole discussion. Companies in the top 1 percent of AI spending are laying out a median of $7,400 per employee. The median company spends $11.95 per employee.

That is not a gap, it is two different economies. AI spending has not diffused across the market; it has concentrated in a very small group of firms, while for everyone else it still amounts to a handful of seat licenses.

Economist Ara Kharazian ties the slow uptake directly to price, arguing that for corporate buyers the extra performance simply is not worth the cost, largely because the return on a marginal capability gain is so difficult to measure.

How much weight should this carry?

Less than a headline would suggest, and a caveat belongs here. This rests on a single source. Ramp's visibility comes from its own customer base, which skews toward US companies using a spend management product, so it is not a representative sample of the global market. First-month figures are also early evidence given how long enterprise procurement cycles run.

What the data does support is a directional claim: the most capable model is no longer automatically the best-selling one. That is a change from where this market sat two years ago.

The version of this we see with clients

A recurring worry in client conversations goes something like this: we are not on the newest model, are we falling behind?

The numbers say no, and that most sophisticated buyers have made the same choice deliberately.

The practical framing is to select the workflow first and the model second. Invoice extraction, email triage, contract summarization and similar work run perfectly well on a mid-tier model, and the accuracy difference typically disappears underneath the process controls you should have anyway. Frontier models earn their price in narrower places: long chains of reasoning, multi-step agent work, large and messy codebases.

Our decision procedure is deliberately boring. Build on the cheaper model, measure the output for two weeks, write down the error rate. Run the identical workload on the expensive model and measure it the same way. If the difference justifies the monthly delta, upgrade. If it does not, stay where you are, because the premium is not a one-time payment. It recurs every month, forever, whether or not anyone is still checking whether it was worth it.

That is the behavior the Ramp figures are actually capturing. Corporate buyers have started doing this arithmetic, and AI has moved from the enthusiasm column of the budget to the line-items-requiring-justification column.

Sources: The Decoder, Ramp AI Index