Article — "Anthropic's Revenue Rockets to $65B Annualized as Its Best Model Loses Users to Cheaper Rivals": Two numbers describe the same company in July, and they point in opposite directions. Anthropic's annualized revenue ran to $65 billion. Its single best model, Claude Fable 5, reportedly accounted for roughly 8% of what customers spent on Anthropic per Ramp's line-item data, though FT-reported figures put Fable 5 at about 11% of total spending.
Hold those together, because the gap between them is the entire story. This is not a company whose flagship is carrying the business. It is a company whose business is being built despite the flagship.
Last week we tracked Anthropic's revenue and profit thread as the run rate climbed from $47 billion in May and the company posted its first adjusted operating profit in Q2. This week the number is $65 billion, confirmed to CNBC and first reported by Bloomberg, and a new dynamic has surfaced underneath it: the premium model that was supposed to justify premium pricing is the one customers are quietly declining to use.
The premium model that was supposed to justify premium pricing is the one customers are quietly declining to use.
Look at Ramp's July breakdown, drawn from billing data across 70,000 companies. The cheaper Opus 4.8 sits at 28.0% of Anthropic spend. Fable 5, still arguably the best model anyone has shipped, sits at 8.0%. And Opus 5, released only on July 24th, is already at 3.5% and climbing.
That last figure is the one to sit with. Anthropic launched Opus 5 explicitly as half the price of Fable 5 while outperforming it on coding and knowledge-work evaluations. Per Ramp's July data cited above, Opus 5 (3.5%) had not yet overtaken Fable 5 (8.0%), though it was climbing fast; some reports reportedly claim it had already overtaken the flagship. Anthropic is not just watching customers route around its best model; it is building the off-ramp itself.
Drew Breunig captured the mechanism better than any market data. For most of the past two years, he writes, "it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems."
Then that pattern broke. "But then Fable landed. It was (and still is!) incredible. But the cost was so high and Opus was good enough (as was 5.6, K3, and even GLM) for most of the code we needed. So we started to think about what work went where."
That last sentence is the whole shift. "What work went where" is procurement language. It means the model is no longer a default; it is a per-task decision, and the frontier model is the exception you escalate to, not the option you reach for first.
The free lunch was the assumption that the newest, best model would also be the cheapest, so choosing it required no thought. Fable 5 ended that. When your best model reportedly costs $10 to $50 per million tokens against rivals under $1, according to cryptobriefing, "best" stops being free, and buyers start doing math they never had to do before.
This is where routing infrastructure stops being a developer convenience and becomes the industry's operating logic. Stripe's agreement to acquire OpenRouter is the clearest signal that the per-task decision is being institutionalized into the plumbing.
OpenRouter offers access to more than 400 models from over 80 providers through a single interface, and it splits two decisions that used to be one: which model handles a request, and which provider endpoint serves it. Requests get evaluated on task complexity, price, speed, and reliability, then directed accordingly.
Read that mechanism against Breunig's "what work went where" and you see the same idea, now automated. The system is designed to send the cheapest adequate model to each task and reserve the expensive one for the cases that genuinely need it. AWS and Snowflake describe their routing the same way: less demanding workloads to smaller, lower-cost models, premium models held back for complex reasoning.
The frontier model, by architecture, becomes the fallback tier.
And this is already the norm, not a forecast. F5's 2026 survey of more than 1,100 IT decision-makers found 52% of organizations chaining or orchestrating multiple models, using an average of seven. Nobody who runs seven models routes everything to the most expensive one.
Ramp's chief economist, Ara Kharazian, framed the reversal with unusual candor. "If you impute previous trends you expect Anthropic to own the market," he said. "But because [OpenAI's newest model] was so good and Fable underperformed, it's been the reverse."
Note the word "underperformed." Fable didn't underperform on capability; by most accounts it is still the best. It underperformed on adoption, which is a different and, for a business, more consequential failure. Being the best and being chosen have come apart.
That decoupling is the strategic core. For years the implicit deal in frontier AI was that state-of-the-art leadership converted to pricing power, which converted to durable margins. Whoever held the crown could charge for it, and enterprises would pay because there was no substitute good enough.
Routing dissolves that logic task by task. When a buyer can send simple work to Opus 4.8 or GLM and escalate only the hard cases to a frontier model, the premium is no longer paid on the whole workload. It is paid only on the sliver that actually requires it. And that sliver, per Ramp, is now single digits.
The consequence for pricing is direct. Premium single-model pricing—the strategy of charging top dollar because you have the best model—cannot survive an environment where being the best only wins you the fraction of traffic that can justify the cost. Every other request goes elsewhere, by design, inside infrastructure that Stripe now wants to own the billing layer for.
Anthropic's own behavior confirms it has already conceded the point. You do not launch a cheaper model that beats your flagship on benchmarks unless you have accepted that the flagship's price is repelling the volume. Average prices paid for Anthropic models reportedly fell nearly 25% since mid-July, according to cryptobriefing. That is not a company defending premium margins; it is a company managing their descent.
None of this contradicts the raw growth. The $65 billion is real, the sevenfold year-over-year climb is real, and 6,000 customers spending $100,000 or more a year is a genuinely large enterprise base. The point is what that revenue is made of: cheaper models, at falling average prices, with the flagship relegated to a fallback role by the buyers' own routing rules.
Which is exactly the question an IPO forces into the open. Anthropic filed confidentially in June, per CNBC, against a valuation reportedly floated at at least $2 trillion for its expected IPO, per KuCoin's report. That number, and the near-trillion-dollar figures already attached to the company, rest on a premise the July data undercuts: that frontier leadership converts to durable, defensible margins.
The July run rate came in below the most optimistic investor forecasts, which had expected the pace to top $80 billion. The miss is not the story on its own. The story is why it might keep happening—because the mechanism steering budgets away from Anthropic's best model is not a competitor's stunt but the industry's own routing architecture, and Anthropic is helping build it.
The crown is still Anthropic's. What's no longer clear is whether the crown pays.