Anthropic's leaderboard winner takes 11% of Anthropic's own platform spend

Ten Anthropic models in Ramp's July table total 63.0 percent of tracked spend ^[1], which means those shares are slices of something wider than Anthropic, and no statement about Anthropic's internal mix survives without renormalising them. Do that and Fable 5 comes to 12.7 percent of the Anthropic spend Ramp can see ^[2], close enough to the 11 percent in the FT's reporting ^[1] that the two are probably measuring the same thing. The rest of the renormalised table is the part a finance team should be shown: 4-series models plus Haiku 4.5 hold 76.0 percent, and everything badged 5 holds 24.0 percent ^[3][4]. Three quarters of the money is sitting on last generation.

The mechanism is arithmetic. Fable 5's list price is twice Opus 4.8's on both input and output ^[8], so it has to return twice the value per task, and the dev.to writeup of the FT data puts the real gain at roughly 1.1x on the workloads it names, summarisation and document extraction among them ^[4]. That is the whole story of the 11 percent.

Where the two accounts split is Opus 5. The dev.to piece has it in the mid-teens and overtaking Fable within weeks of its July 24 launch ^[6]. Ramp's July table has it at 3.5 percent against Fable's 8.0 ^[7], which renormalises to 5.6 percent of Anthropic spend ^[5]. Simon Willison supplies the reason for the gap himself: July contained one week of Opus 5 ^[9]. So the cannibalisation claim is a forecast rather than a reading, and CNBC's report that Opus 5 beats Fable 5 on coding and knowledge-work evaluations at half the price ^[14] is the case for expecting it to come true, not evidence that it has.

One number in the dev.to piece does not survive a check. It calls the gap between Fable 5 output and DeepSeek V4-Flash output 71x ^[13]; the two prices it prints, $50 and $0.28 per million tokens ^[3][12], are 179x apart ^[6]. The 71x figure is what you get from the discounted $20 output price on OpenAI's Sol measured against the same floor ^[11][7]. The direction of travel is the same either way. A price table that mislabels its own ratio by a factor of two and a half ^[9] is still not the slide to paste into a procurement deck.

None of this is a demand problem. Anthropic's annualized revenue reached $65 billion in July, up from $47 billion in May, with 6,000 accounts spending $100,000 a year or more ^[10]. It is a mix problem, and on the buyer side mix is usually one config value. The default model in a gateway gets set once, by whoever wired it up, and on these prices that line carries a factor of two on every token that passes through it. It is rarely owned by anyone who reads the invoice.