Ramp went looking for the money and didn't find it. In Fable 5's first month on general sale, businesses bought six percent of their Anthropic tokens from it and spent 11.4 percent of their Anthropic dollars there. The gap between those two numbers is the price doing its work: at $10 per million input tokens and $50 on output, Fable runs double Anthropic's other flagship tier, so a thin slice of usage swells into a fatter slice of the bill. Even swollen, it's thin.

The cross-vendor figure is the one that stings. Fable generated roughly three-quarters as much model-attributed spend in July as OpenAI's GPT-5.6 Sol, while charging about twice as much per token. Sol also took a far larger share of its own house's usage, 25 percent of OpenAI tokens against Fable's six, though those are shares of two different totals and shouldn't be read as a like-for-like volume comparison. Ramp's own caveat matters more anyway: its sample of 70,000 businesses skews tech-heavy, so adoption across the wider economy is probably thinner than six percent rather than fatter.

Price is the easy read and it's mostly the right one. Anthropic shipped Opus 5 on 24 July at $5 and $25 per million tokens, half of Fable on both sides, and said it beat Fable on coding and knowledge-work evaluations while being "designed to be used every day." That is a company telling its customers, in the politest available language, that the flagship isn't the default. The best public cost comparison I can find doesn't quite test the right thing: the consultancy ML6 ran Fable against Opus 4.8, a generation back, and found Fable caught release-blocking issues that 4.8 missed at about three times the wall-clock and 3.4 times the money, $12.38 against $3.65. Against Opus 5 the price premium would be narrower and the capability gap narrower still. That is the comparison every buyer has actually been running since late July, and it's the one nobody has published.

The consumer side ran the same argument faster and in public. Anthropic bundled Fable into subscription limits, announced a move to metered usage credits on 7 July, pushed the deadline to the 12th, pushed it again to the 19th, and then on 20 July put Fable back into Max and Team Premium at half of plan limits while leaving Pro on credits with a one-time hundred-dollar sweetener. Three schedules in thirteen days is a company finding out in real time what its subscribers will tolerate paying for the best thing it makes.

None of which is visibly hurting Anthropic. It recorded its first adjusted operating profit in the second quarter, guided investors toward another in the third, and reached roughly $65 billion in annualised revenue by July. One tier underperforming is a product problem, not a solvency one.

Which is where phones come in. The average American handset was 3.16 years old when it got traded in during 2020, and by 2025 that had stretched to 3.84 years, with the global cycle walking up from 2.4 years in 2013 to 3.7 by 2022. A longer cycle on its own doesn't prove that new phones stopped feeling better, so the more useful number is the reason people give when they finally do upgrade: three quarters of them say it was the battery. Not the camera, the processor, or anything a keynote gets built around. Once the dominant trigger for replacing a device is a physical component wearing out, the feature improvements have already stopped doing the work of selling it. They justify the purchase after the wear has forced it.

Ramp's economist reads the Fable numbers as a ceiling, saying the company has "found a new upper bound for how much businesses are willing to spend on AI." I think that's slightly the wrong shape. The data shows a pricier, more capable model taken up more slowly than a cheaper one. It doesn't establish a budget cap, because total enterprise AI spending kept climbing through the same period and the share of Ramp businesses paying for AI at all went from just over half in March to nearly 56 percent by July. Buyers will pay for a gain they can see in a number they already keep. Eleven percentage points on SWE-Bench Pro is a real difference and it is invisible in every metric a finance team tracks, which is the same trap the mini tiers exposed back in March.

The analogy has a limit, and working through it changes the answer rather than softening it. A phone is one visible purchase every few years and the comparison is trivially easy: this handset, that handset, this price. Tokens are metered continuously, and the choice gets made per task by an engineer who mostly never sees the invoice, which ought to make reaching for the expensive model easier rather than harder. Except that Fable carries friction no handset does. Anthropic set the safety classifiers deliberately wide, writing that it configured them "to trigger on a set of requests that we know are likely benign" with a margin "much larger than in any prior launch," and that users experience this as the model refusing reasonable, non-harmful requests. Sitting on top of that is a mandatory thirty-day data retention policy with no configuration toggle and no enterprise carve-out. The engineer isn't unconstrained after all: one of those constraints lives in the code path and the other lives in the contract.

So the six percent is overdetermined, and I'd rather say that than pick the tidier cause. Price, refusals and retention all push the same direction, and this data can't cleanly separate them. What the phone comparison adds is the part that survives all three: strip the frictions out entirely, and a capability gain nobody can feel still doesn't move a purchase, which is roughly what the NBER found when a flood of new software moved no usage at all.

There is an AI equivalent of a dying battery, and the labs own the schedule for it. Nothing wears out in a model. The thing that eventually forces the upgrade is the old tier being retired, deprecated, or quietly repriced out from under the people who built against it. On the current numbers that lever moves more adoption than shipping something better does, which is an odd place for a research company to arrive.

Sources: