On Wednesday August 12, 2026, Ramp Lead Economist Ara Kharazian published the August AI Index under the headline "Cracks in the AI thesis." Buried in the model-share section is the sentence the whole report exists to publish. In its first thirty days on the market, Anthropic's Fable 5 — the model Anthropic itself has repeatedly described as the strongest general-purpose model shipped by any lab in 2026 — captured six percent of the tokens businesses bought from Anthropic. The comparable OpenAI number, for GPT-5.6 Sol, is twenty-five percent. The Decoder picked it up within hours, and Implicator ran the numbers against Anthropic's rumoured $2 trillion IPO valuation the same afternoon.

Fable 5 is roughly twice as expensive per token as GPT-5.6 Sol at $10 per million input and $50 per million output. Its share of Anthropic dollar spend is therefore higher than its share of tokens — 11.4 percent — but nowhere near the twenty-three percent GPT-5.6 Sol pulled on the OpenAI side. Ramp's July run of Fable 5 revenue came in at 75 percent of GPT-5.6 Sol's, on a model that had been available for thirty days and is, by benchmark, the more capable product. Kharazian's summary sentence, on the record: "So with Fable 5, we've found a new upper bound for how much businesses are willing to spend on AI."

Anthropic is not losing. It is winning on the wrong axis.

Anthropic still leads the vendor count

The Loop covered May's Ramp crossover — 34.4 percent Anthropic against OpenAI's 32.3 percent, the first month Claude had led the index. Three months on, the vendor-count gap has widened. 43.5 percent of US businesses on Ramp's panel now use Anthropic, up 1.1 points month-on-month. 39.7 percent use OpenAI, up 0.23 points. xAI is at 4 percent, up 0.94 points — Ramp calls it "the fastest growth since July 2025." The bucket Ramp labels "model serving platforms" — the one that carries open-weights traffic and Chinese-lab APIs — climbed to 6.1 percent, up 0.2 points.

Read the two panels side by side and the story bends. Anthropic is adding paying customers faster than any other frontier vendor. Those customers are not buying Anthropic's flagship. The bucket that carries open weights is growing faster than the bucket that carries Fable 5.

What Anthropic's own July repricing already implied

The August 12 number is not a surprise inside the building. On July 24, Anthropic launched Opus 5 at $5/$25 per million tokens — the same price as Opus 4.8 before it, but described in the launch note as "frontier intelligence at half the price of Fable 5." That single sentence repositioned Fable 5 from the flagship a customer buys to the flagship a customer buys when Opus 5 is not close enough. Six days earlier, Anthropic had already stripped Fable 5 out of the Pro and Team Standard subscriptions and pushed those users to API rates. When the vendor is pricing its top-tier model against its own second-tier model in the launch copy, the vendor already knows what Ramp is going to measure the next month.

Kharazian's "upper bound" line is what the July repricing operation looks like once the buyers vote with their corporate cards.

The AWS CTO said the quiet part in July

Werner Vogels, Amazon's CTO, told Fortune on July 10 that customers were asking whether they really needed the biggest model available. His verbatim answer: "Do you really need the biggest, highest-end model to solve this? The answer is no, you don't." Read Vogels in the same paragraph as Kharazian and the shape of the enterprise conversation is clear. The CTO of the world's largest cloud reseller is on record telling customers to buy something smaller. The economist tracking 50,000 US companies' expense cards has now published the number that shows they took the advice. And the flagship's own vendor, in the same six weeks, halved its own launch reference price and pushed Pro subscribers to the API.

Three data points, one story. Frontier pricing has a ceiling and the ceiling sits below the headline rate on the largest models in the market.

The per-employee number nobody wants to publish

The other statistic in Kharazian's post that is worth reading twice concerns per-employee spend. The top one percent of businesses spend $7,400 per employee per year on AI. The top ten percent spend $650. The median company spends $11.95 — eleven dollars and ninety-five cents, per employee, per year. The distribution is that fat-tailed. Anthropic's Series H and OpenAI's Deployment Company are being underwritten by the top one percent, and the top one percent has now shown that even at their spending altitude, the marginal call on the most expensive model on the market is no.

That distribution is the sound of an AI market maturing on the wrong beat for the labs. The customer base grew. The share going to open weights grew. The share going to the frontier flagship shrank. When the top of the willingness-to-pay curve hits a ceiling and the bottom of the market is spending less than the price of a supermarket sandwich, the growth-equity thesis has to be rewritten in something more honest than a launch tweet.

Ramp's methodology note is honest enough to add that the panel "skews slightly more tech-y than our typical AI Index sample," meaning the six-percent Fable share is a ceiling on adoption in the tech-forward tail of US corporate spend. Across the broader corporate universe, the number is lower still.

What to watch

1. Whether September's Ramp Index shows Fable 5 recovering. The August 12 report covered July. Anthropic will spend August running its Fable 5 marketing against the six-percent number, and the September Index will show whether the Opus 5 repositioning, the Max repackaging, and the API push have moved enterprise buyers back to the top tier. If Fable's token share stays below ten percent for two straight months, Anthropic will have to decide whether Fable 5 is a marketing surface or a revenue surface.
2. Whether Ramp adds a Google line. Gemini 3.7 Flash launched on August 13 at half the price of its three-week-old predecessor. The August Index does not break Google out as a separate vendor row, and Gemini's pricing pressure on GPT-5.6 Sol and Opus 5 is not visible in Kharazian's numbers. If the September issue carries a Google column, the frontier-vendor share picture changes for a third time in a quarter.
3. Whether the Anthropic IPO paperwork prices the ceiling. Implicator's August 13 write-up says six Anthropic backers are modelling annualised revenue of $100–120 billion by year-end ahead of a rumoured $2 trillion IPO. Those are extrapolations, not booked revenue. If the S-1 disclosures — when they land — carry the same Ramp-shaped ratio between adopter count and flagship-model spend, the market will get to price the ceiling directly.
4. Whether "upper bound" gets said by a hyperscaler CFO. Kharazian works for a corporate-card company. When the same sentence appears on a Microsoft, Amazon, or Google earnings call — the one where a hyperscaler CFO tells the analysts on the line that their AI customers are trading down — the ceiling stops being a Ramp datapoint and starts being a public-market fact.

The single most quietly striking line in the August 12 report is not the six-percent number. It is the sentence next to it, admitting the sample skews "slightly more tech-y" than the average American company. If the tech-forward tail of US business buys the most expensive AI model on the market six percent of the time, the rest of American business is buying it less. Kharazian's methodology paragraph is what honesty looks like in an AI Index. Anthropic's July repricing is what honesty looks like in a launch note. The market, in between, has been honest enough to buy something else.

The Fable 5 launch called it "the best model ever shipped." One month in, six percent of Anthropic's own token volume agrees.