There is a particular kind of silence that settles over a room when someone confesses they paid double for something and then quietly stopped using it. It happened to me last week during a governance call, when a protocol strategist admitted that his team had migrated every workflow off their premium AI model within three weeks of deployment. "It was like buying a Ferrari," he said, "to drive to a grocery store two blocks away."
That anecdote came rushing back when I read the data from Ramp, the corporate spend management platform. Anthropic's flagship Fable 5, the model that was supposed to define the next frontier of frontier intelligence, accounts for only 11.4% of the company's enterprise spending. Not tokens. Not deployments. Spending. In the world of AI procurement, that number is not a footnote. It is a verdict.
When Capability Meets the Real World
Let me give you the lay of the land. Fable 5 is Anthropic's most powerful model to date. It prices at $10 per million input tokens and $50 per million output tokens โ exactly double the price of Opus 5, the model that sits just below it on the pedestal. And that's the entire story, really. Anthropic themselves positioned Opus 5 as "frontier intelligence that approaches Fable 5, at half the price."
I've seen this before. In 2017, when I was drafting the Polymath whitepaper on tokenized equity, I spent weeks negotiating with legal counsel over a single clause about voter rights. We made it philosophically beautiful and utterly impractical to implement. It took three months to untangle that mistake. The same pattern plays out here: a team so enamored with what they could build that they forgot to ask whether anyone would actually want it.
Over the past 30 days, the enterprise AI market has effectively sorted itself into two camps. On the one side, you have the model that makes you feel the future. On the other, the model that makes your quarterly budget look defensible. The Ramp data suggests the budget is winning. Fable 5 accounts for just 6% of Anthropic's total token usage, despite carrying 11.4% of the spend weight. That gap โ that awkward spread between usage and dollars โ tells me that the handful of companies using Fable 5 are doing so for narrow, high-stakes tasks: complex reasoning chains, long-horizon planning, document analysis that requires absolute precision. And everyone else has moved on to Opus 5.
The Diminishing Returns of Smarter
Here's the insight that matters, and it's not the one Anthropic wants to sell you. The performance gap between Fable 5 and Opus 5 is far smaller than the price gap. In my own testing โ I ran a battery of long-form governance document analyses through both models last month โ the difference is noticeable but not transformative. It's the difference between 90 and 95 on an exam. The absolute jump in capability exists, but the marginal utility for most enterprise workflows approaches zero.
This is the diminishing returns curve, and it's a brutal mistress. Getting from 80 to 90 costs a certain amount of compute. Getting from 90 to 95 costs ten times as much. And yet, the customer experience barely shifts. Fable 5 is the model that makes you say, "That's nice," while Opus 5 is the model that makes you say, "That's enough." And "enough" is now the most powerful word in enterprise AI.
Accel partner Miles Clements put it plainly when I asked him about the data: most people don't need to use frontier models continuously. He's right. The enterprise has discovered that the overwhelming majority of business workloads โ writing SQL queries, summarizing meeting notes, drafting client emails โ live in the "good enough" zone. The margin between Opus 5 and Fable 5 on these tasks is invisible. The margin in price is not.
But what interests me most is what this says about the market's emotional structure. We are seeing the death of the "best model" narrative as a commercial driver. It's not that businesses don't want intelligence. It's that they now understand that intelligence has a gradient, and most of their tasks live on the cheap end of that curve.
The Contrarian Blind Spot
Now, let me argue against myself for a moment, because this is where the story gets uncomfortable. The Ramp data might be measuring the wrong thing. Fable 5 is only two months old. The enterprise procurement cycles โ I've been in them, I know how they drag โ mean that many serious adopters are still in evaluation mode. A two-month window is not a verdict. It's an early draft.
I remember the first month of the MakerDAO governance work in 2020. We saw a low early participation rate and nearly wrote off the entire risk parameter overhaul as a failure. Then the community woke up, and by the third month, we had a 70% participation rate. Adoption curves are slow at the base. It takes time for trust to build, for workflows to be reimagined, for security reviews to clear.
And there's another blind spot. The Ramp data only covers companies that use Ramp. It is a sample, not the universe. The biggest enterprises โ the ones with the deepest pockets and the most complex needs โ might be buying Fable 5 through private contracts and direct deals that never show up in these numbers. We might be looking at a shadow of the true demand.
But here's the counter-counterpoint that makes me believe the initial data matters: Opus 5 was released at the same time, and it's already overtaken Fable 5 in a shorter window. If enterprises were slow to adopt Fable 5 because of evaluation cycles, why were they so fast to adopt Opus 5? The answer is simple. They didn't need to evaluate. The value proposition was obvious: near-frontier intelligence at half the price. That is the easiest purchase decision in enterprise AI history. Fable 5's evaluation process is ongoing because the price demands justification. And that justification is a hard ask.
The Value Layer Has Arrived
I have lived through enough cycles to recognize a structural shift when it shows up in the numbers. And this is one. We are not witnessing a temporary pricing correction. We are witnessing the maturation of AI commercialization โ the moment when the market stops buying the best and starts buying the optimal.
Fable 5 is the first model to be the "best" and simultaneously be rejected. That's a new thing. We didn't see this with GPT-4. We didn't see this with Claude 3 Opus. We saw the market line up for the most powerful tool. But now, for the first time, the market has said: we see what you are, and it's not worth it.
I was talking to a founder last week about this, and she said something that stuck with me: "We used to chase the SOTA. Now we chase the ROA." Return on ability. The number of tasks you can complete per dollar. That is the new metric, and it doesn't favor the flagship.
This is good news, actually. It means the AI market is becoming rational. It means companies are going to stop buying bragging rights and start buying outcomes. It means the model routing layer โ the software that decides which model to send each task to โ is about to become the most valuable infrastructure in the stack. It means Anthropic, despite Fable 5's stumble, is well-positioned because Opus 5 is a genuine hit.
What's less clear is whether Anthropic will make the same mistake again. Will they launch Fable 6 with the same premium price and hope for a different result? Or will they internalize the lesson that the market has now taught them at scale: that being the best is the worst business model in AI. The margin of "the best" is too thin, the price elasticity too steep, and the patience of enterprise customers too short.
The path forward isn't to make models smarter. It's to make the smart the models available at prices that make sense. We are entering the era of the model as a utility, not as a luxury. And the utility model has never needed to be the best. It only needs to be good enough, at the right price, with the right trust. The soul of this industry is shifting, from the chase of the frontier to the curation of value. And the silent vote of the enterprise wallet is the loudest signal we have.
We stand at a moment where the question is no longer "what can these models do?" but "what will we choose to pay for?" And the answer, from the data, is a quiet: enough.