Fable 5 Is 6% of Anthropic's Tokens. Businesses Found Their Ceiling.

TL;DR

Ramp, which sees the card and bill-pay spend of more than 70,000 US businesses, published its August AI Index with the first public uptake data on Claude Fable 5. Over the month Ramp measured, Fable 5 was 6% of the tokens businesses bought from Anthropic and 11.4% of the dollars, despite being the best and most expensive model on the market. OpenAI's flagship, GPT-5.6 Sol, was 25% of OpenAI tokens at half the price. The Financial Times ran the data on August 23 with a fresh detail, that the cheaper Claude Opus 5 has already passed Fable 5 in enterprise spending, and the story spent Sunday on the Hacker News front page with over 300 points. Ramp's economist calls it "a new upper bound for how much businesses are willing to spend on AI."

What Ramp actually measured

The index is not a survey. Ramp's lead economist, Ara Kharazian, pulls it from what companies pay for through Ramp cards and bill pay, and the model-level cut comes from Ramp's token spend management product, which tracks daily usage by model. The August edition, dated August 12, covers July, the month after Fable 5 came back from its June export-control pause.

The headline paragraph is blunt: "Over the last month, Fable 5 has made up only 6% of tokens businesses purchased from Anthropic, and despite being their most expensive model by far, 11.4% of dollars spent on Anthropic models." For comparison, GPT-5.6 Sol "comprises 25% of OpenAI tokens and 23% of spend," and Fable 5 generated roughly 75% as much model-attributed spend as Sol did. Kharazian's verdict: "Fable 5 is less popular with businesses than GPT-5.6 Sol overall."

One caveat Ramp puts in its own post: the token-spend sample "skews slightly more tech-y" than the wider index, which means real-world Fable adoption across all businesses is probably lower than 11.4%, not higher.

The price card is the whole story

Anthropic's price card makes the gap obvious. Fable 5 is $10 per million input tokens and $50 per million output. Opus 5, released July 24, sits at $5 and $25, on the same rate as Opus 4.8. Sonnet 5 is $2 and $10, a price Anthropic has now made permanent after the planned September increase was cancelled. Haiku 4.5 is $1 and $5. Sol, after OpenAI's August 21 cut, is $4 and $20.

Kharazian's framing: "Fable 5 is the most performant model on the market. It's also the most expensive, at roughly $10 per 1M tokens, twice as expensive as the still highly performant GPT-5.6 Sol." Then the sentence that gave the report its title, "Cracks in the AI thesis": "So 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."

There is a wry footnote on that same price card. Anthropic will happily sell you Opus 5 at Fable 5 prices, $10 in and $50 out, if you turn on Fast mode. Which is one way to find out whether the ceiling is about the price or the model.

What customers bought instead

Ramp's interactive index also charts model-attributed spend by model across every provider it tracks. For July, the single biggest line item in that pool is Opus 4.8, at 29.0%. Fable 5 sits at 6.2% of the pool. Opus 5, with one week on sale, was already at 3.3%. That is a different cut from the 6% and 11.4% in the blog post (different denominator, different sample), so do not stack them, but the shape is the same: the workhorse is the previous Opus, and the flagship is a sliver.

The Financial Times extended the picture on August 23. Per its report, relayed by PYMNTS and others, Fable 5 has settled at about 11% of what customers spend on Anthropic more than two months after launch, and Opus 5, launched in late July at half the price, has already overtaken it in enterprise spending. Accel's Miles Clements gave the FT the line that will get quoted back at Anthropic for a while: "Most people don't need to operate at the frontier," and the era of customers picking only frontier models "was not a durable era."

If you want an analogy: a hospital's most expensive specialist sees a small fraction of patients, and the building is designed so that fraction is enough. Fable 5 is the specialist. The waiting room runs on Opus 4.8 and Sonnet 5, and the routing desk is now a product category. Kharazian's August 19 vendor rundown notes that businesses are "increasingly using routers and model serving platforms to access cheaper models outside of those offered by OpenAI and Anthropic," with five router or model-serving vendors on Ramp's trending list, and OpenRouter acquired by Stripe.

The retention tax

Price is not the only drag Ramp names. Kharazian told Quartz this week: "GPT-5.6 Sol is really good, increasingly the choice for developers. Fable 5, meanwhile, disappointed both in adoption and real-world application given price + data retention requirements imposed by regulators." The FT write-ups repeat the point, attributing to Kharazian that data retention rules introduced under the Trump administration held Fable 5 back.

The context, which we have covered before: since mid-June Anthropic retains 30 days of traffic on its covered frontier models, zero-data-retention contract or not, while OpenAI has said it will keep offering ZDR on its frontier tier. For a bank, a law firm, or anyone with a customer data agreement that says "no retention," the best model on the market is one their compliance team cannot sign off on. One of the top Hacker News comments on the FT story calls the missing ZDR option "the largest barrier" to deploying Fable in an enterprise, and another describes dropping the $200 Claude plan for Codex after Fable kept hitting usage limits mid-session.

Anthropic is still winning the vendor race

None of this means Anthropic is losing customers. The same August index has Anthropic at 43.5% of US businesses paying for AI, up 1.1 points on the month, against OpenAI's 39.7%, up 0.23. xAI is at 4%, up 0.94, its biggest monthly gain this year. Anthropic passed OpenAI in paid adoption for the first time in May and has widened the gap every month since.

The wider index tells a quieter story underneath. Ramp's overall adoption rate, the share of US businesses paying for any AI product, reached 55.7% in July, up from 47.0% in January. But the monthly gains have shrunk from 2.2 points in March to 0.8 points in each of June and July. Spending is also lopsided: the top 1% of businesses spend a median $7,400 per employee per month on AI, the top 10% spend $650, and the median firm spends $11.95. The flagship-model question is really a question about that top 1%, because they are the only buyers for whom a $50 output token was ever going to be routine.

What the ceiling means if you build on these APIs

• Routing is now table stakes, not optimization. The revealed preference in Ramp's data is that the top model handles a minority of tokens. If your product hard-codes the flagship for every call, you are pricing yourself above your customers' own behavior.
• Half price at near-parity is the winning slot. Opus 5 and Sol both sit at roughly half of Fable 5's rate and both are taking share. Anthropic's own pricing signals the same read: the 2x tier is for Fast mode and Fable, everything else is $5 or less on input.
• Compliance is a feature with a price. ZDR, or the lack of it, is showing up in spend data now, not just in procurement threads. If you sell into regulated customers, the model with the retention carve-out is the model you can ship.
• The IPO math has to route around this too. Anthropic's shareholders are reportedly anticipating an October IPO above $2 trillion. Ramp's data says the growth driver is breadth, 43.5% of businesses and climbing, not the premium tier. That is a fine story, just a different one from "everyone upgrades to the frontier."

Caveats

Ramp's sample is Ramp's customers: US businesses that pay for AI through its cards and bill pay, which is not a random slice of the economy. The model-level figures come from a smaller, more technical subsample, and Ramp says so. Token shares depend on tokenizers: Anthropic's 4.7-and-later models, Fable included, produce roughly 30% more tokens for the same text than Sonnet 4.6 and earlier, which if anything inflates Fable's token share against the older models still in the mix. The FT article is paywalled; its Opus 5 claim and the Accel quote are taken from multiple independent write-ups that agree with each other, but we could not read the original. And Fable 5's July is not a clean month: it was pulled from sale in mid-June by a US export-control order about 72 hours after launch, returned to subscriptions permanently on July 20, and is capped at 50% of limits on Max and Team Premium plans, all of which suppresses usage in ways that have nothing to do with willingness to pay.

Key Takeaways

• Ramp's August AI Index puts Claude Fable 5 at 6% of the tokens and 11.4% of the dollars US businesses spent with Anthropic over the month it measured, versus 25% and 23% for GPT-5.6 Sol at OpenAI.
• Fable 5 lists at $10 in / $50 out per million tokens, twice Opus 5 and Opus 4.8, 2.5x Sol after its August 21 cut, and 5x Sonnet 5.
• The Financial Times reports Fable 5 has settled near 11% of Anthropic spend two months after launch and that Opus 5, launched in late July, has already overtaken it.
• Ramp's economist blames price plus regulator-driven data retention requirements; Anthropic retains 30 days of covered-model traffic even under ZDR contracts, OpenAI does not.
• Anthropic still leads paid business adoption, 43.5% to OpenAI's 39.7%, and has widened its lead every month since passing OpenAI in May.
• Overall US business AI adoption hit 55.7% in July, but monthly growth has halved since March, and the median firm spends $11.95 per employee.

Sources: Ramp AI Index, August 2026 (Cracks in the AI thesis), Ramp Economics Lab mirror of the same post, Ramp AI Index interactive data, Ramp AI Index, July 2026, Ramp AI Index, June 2026, Kharazian, Top SaaS vendors on Ramp, August 19, Anthropic pricing, OpenAI on Zero Data Retention, Quartz, August 21, PYMNTS on the FT report, Hacker News discussion of the FT story.