Anthropic markets Fable 5 as its most capable model yet. Enterprise buyers, judging by new spending data, are not rushing to pay for it. In its first month on the market, Fable 5 made up roughly six percent of token volume purchased through Anthropic’s API, according to financial services firm Ramp, and 11.4 percent of dollars spent on Anthropic models overall.

That gap between usage share and revenue share traces back to price. Fable 5’s list price, about $10 for a million input tokens and $50 for a million output tokens, runs roughly double what Anthropic charges for its other flagship models and well above OpenAI’s GPT-5.6 Sol.

The comparison to OpenAI is unflattering for Anthropic’s flagship. GPT-5.6 Sol accounts for 25 percent of OpenAI’s token volume and 23 percent of its API spending, and generated about a third more model-related revenue than Fable 5 over the same stretch, despite its lower per-token price. Ramp’s sample draws from its own spend-management product and skews toward tech companies; the firm notes actual Fable 5 uptake across the broader economy is likely lower still, since the model is used mainly for coding work.

Ramp economist Ara Kharazian reads the pattern as evidence that corporate willingness to pay for frontier-tier AI has found a ceiling. That is Kharazian’s interpretation of the data, not something the numbers establish on their own. What Ramp actually measured is that a pricier, more capable model is being adopted more slowly than a cheaper rival. Whether that reflects a hard budget cap, or simply that Fable 5’s performance edge is difficult to measure in ordinary daily work and therefore hard to justify paying for, is a separate question the spend data alone cannot answer. A model that produced a clearly measurable jump in output could still command a premium. The constraint appears to fall on paying for gains that stay abstract, not on frontier pricing as such.

The timing sharpens the question. DeepSeek released a new model this week priced under a dollar per million output tokens and claiming benchmark wins over Anthropic’s Opus 4.8, and Microsoft has been substituting its own cheaper in-house models into Copilot and other products instead of defaulting to third-party frontier models. Neither development proves Kharazian’s ceiling thesis. Both give enterprise buyers a cheaper option to test against Fable 5 before renewing at the higher price.

Growth across the frontier-model market is slowing more broadly, per Ramp:

Ramp’s data also shows open-source models closing the capability gap with frontier releases to within a few months, which the firm links to slowing growth at both leading labs as advanced users shift spending toward cheaper open alternatives.

Total enterprise AI spending is still climbing, though. Ramp put July’s median AI spend per employee at $7,400 for the priciest one percent of companies, $650 for the top decile, and $11.95 for a typical company. Spending is not contracting. It is concentrating on models whose value is easiest to verify.

Any team renewing a frontier-tier contract this quarter should run its own workload, not a vendor’s benchmark slide, through both a premium and a mid-tier model before defaulting to the most expensive option. If the quality difference doesn’t show up in the metric the business already tracks, the cheaper model is the correct default until it does.

The Decoder reported this analysis of Ramp’s enterprise spending data on August 13, 2026.