Usage of Anthropic's flagship model, Claude Fable 5, has been more limited than the company's other models. According to data published by Ramp on August 12, 2026 for the month of July, Fable 5 accounted for just 6% of the tokens that companies using its token management product purchased from Anthropic. While it accounted for 11.4% of spend, this was still less than the model-attributed spend for OpenAI's flagship model, GPT-5.6 Sol. AI spending itself is rising, so we need to separate whether the low usage ratio stems from pricing or from the evaluation periods and use cases of the companies adopting it.
The Gap Between 6% of Tokens Purchased and 11.4% of Spend
According to Ramp Economics Lab, Fable 5 accounted for 6% of the tokens companies purchased from Anthropic over the most recent month and 11.4% of spend on Anthropic's models. By comparison, GPT-5.6 Sol accounted for 25% of tokens and 23% of spend within OpenAI. Comparing July's model-attributed spend directly, Fable 5 came to about 75% of GPT-5.6 Sol's figure.
This gap doesn't mean Fable 5 isn't being used. The fact that its spend share exceeds its token share indicates that some companies are paying a premium price to use it. Still, even for Anthropic's top-tier model—positioned for long-duration coding and complex knowledge work—usage remained confined to a portion of the company's overall footprint. This ratio could arise either from deliberate restriction to high-difficulty cases or from adoption that has not yet spread company-wide.
Ramp's Ara Kharazian interpreted this result as a new ceiling on what enterprises are willing to pay for AI. The view is that higher performance doesn't automatically justify a higher price. However, what can actually be confirmed from public data is only the relationship between price and usage—it does not pin down price as the specific cause of the low usage ratio.
How Much Work Inside a Company Justifies Paying 2x the Price?
Fable 5's API pricing is $10 per million input tokens and $50 per million output tokens. Claude Opus 4.8's standard pricing is $5 and $25 respectively—meaning Fable 5 costs exactly twice as much for both input and output. If success rates improve on large-scale code migrations or complex research tasks, that price difference can be recouped. But if the same model is also used for summarization, classification, and routine code generation, the bill balloons faster than the value generated by the performance gap.
In enterprise procurement decisions, what matters is total cost per task, not benchmark rankings. If a model reaches the correct answer using fewer tokens, the total cost can be lower even at a higher unit price. Conversely, if retries and human verification don't decrease, performance gains can't be converted into cost savings. Ramp's aggregated data doesn't include task-specific success rates, output volume, or retry counts, so it's impossible to separate whether Fable 5 was overpriced or deliberately limited to high-difficulty cases.
The price sheet itself contains conditions that shift the effective unit cost. Anthropic offers a 90% discount on Fable 5's input tokens when prompt caching is used. US-only inference costs 1.1x the standard rate for both input and output. Workloads that reuse a fixed, long context behave very differently in terms of profitability from workloads that generate large amounts of varying output each time—even on the same model—so list price alone can't determine whether adoption makes sense.
AI Spending Is Growing, But Not Concentrating on the Top-Tier Model
In July, the share of US companies paying Anthropic reached 43.5% according to Ramp's AI Index, up 1.1 points from the prior month. OpenAI rose 0.23 points to 39.7%, and xAI rose 0.94 points to 4.0%. Since Anthropic's customer base is growing, Fable 5's low share of the mix shouldn't be read as declining demand for Anthropic's products overall.
Spending isn't shrinking either. In July, median AI spend per employee was $7,400 for the top 1% of spenders, $650 for the top 10%, and $11.95 for the median company. While leading companies are investing heavily, that money isn't flowing in one direction toward the most expensive model. It's easier to manage budgets by using cheaper models for routine company-wide tasks and reserving the top-tier model only for the small number of cases where failure costs are high.
Pathways to cheaper models are also starting to widen. Among AI-using companies, 6.1% use model-serving platforms that provide access to open-source models and some Chinese-made models—up 0.2 points from the prior month. However, Ramp itself flags this figure as an imperfect proxy. As of June, 96.4% of companies using serving platforms also used at least one of OpenAI or Anthropic. What can be confirmed is only that companies use both together—public data doesn't reveal how companies allocate models on a per-task basis.
Two Caveats Behind the "First Month on Sale" Numbers
Fable 5's one-month data has a different timeline than a typical new product launch. Anthropic released it publicly on June 9, but halted access for all customers on June 12 in response to a US government export control directive. It resumed global availability on July 1 after the restriction was lifted, and through July 7 counted usage under Pro and Max weekly usage limits at up to 50%. The same terms applied to Team and some Enterprise plans before shifting to usage credits afterward. This month's figures measure July after the relaunch, rather than a continuous first month on sale.
The sample size is also limited. Fable 5's token ratio wasn't calculated directly from Ramp's overall transaction data, which covers over $100 billion in annual spend and more than 50,000 US companies. Rather, it's drawn from companies using the AI Token Spend Management product, which tracks daily token usage—a sample that skews somewhat more toward tech companies than the standard AI Index. The number of companies included, contract discounts, and cache utilization rates have not been disclosed.
Even so, the 6% figure leaves model companies with clear homework to do. To convert differences on a performance chart into revenue, they need to show how many tasks succeeded when handed to the top-tier model, how many hours of human labor were saved, and how much lower the total cost was compared to cheaper models. If retention and task-specific spend don't grow over the coming months, Fable 5 won't represent an absolute ceiling on what companies are willing to pay—it will instead mark the price band that a model can't cross unless its performance advantage can be translated into dollar terms.