Anthropic's Priciest Model Wins 8% of Its Own Spend • The Weighted Average

Anthropic’s most expensive model captured 8.0% of US business spend on AI models in July, while a cheaper model released two months earlier held 28.0%, according to card-transaction data in the Ramp AI Index. The company’s revenue is not the problem — the run rate reached $65 billion — but the mix inside that revenue is telling buyers something the price list does not: the market has decided which Claude it wants, and it is not the flagship.

The arithmetic that matters for a procurement team is simple. For every dollar businesses spent on Opus 4.8, a May-era model, they spent 41 cents across both of Anthropic’s July releases combined — Fable 5 at 8.0% and Opus 5 at 3.49% against Opus 4.8’s 28.0%. A vendor whose newest and priciest tier cannot displace its own predecessor is a vendor with pricing power in its investor deck and none in its customers’ invoices.

The flagship nobody buys

Start with what the two models cost. On the Claude pricing page, Fable 5 lists at $10 per million input tokens and $50 per million output tokens; Opus 5 lists at $5 and $25 — exactly half. Anthropic said as much when it shipped the cheaper model, describing Opus 5 in its Claude Opus 5 announcement as coming “close to the frontier intelligence of Claude Fable 5 at half the price,” and naming it the new default on Claude Max. When a vendor markets its second-best model as nearly-as-good for half, buyers are not being subtle in taking the offer.

They have not fully taken it, either. Opus 5 launched on July 24 and reached only 3.49% of tracked spend by the end of the month, a partial month at best. The share that actually moved went nowhere new: Opus 4.8, released in the spring, still absorbed 28.0% of all model spend in Ramp’s July panel, more than Fable 5 and Opus 5 together by a factor of nearly two and a half.

Anthropic's premium flagship stalled where its cheaper model surged

Share of US business AI model spend, monthly, with list price per 1M output tokens

Ramp AI Index; Anthropic pricing page · Jul 2026

The adoption curves make the contrast visceral. Opus 4.8 went from 1.62% of spend in May to 30.95% in June — a full generational takeover inside four weeks — before settling at 28.0% in July. Fable 5, over its own first two months in the same data, moved from 1.82% in June to 8.0% in July. Same vendor, same buyers, same window, two very different verdicts. The variable that separates them is the number on the invoice.

None of this reads as distress at the top line. Anthropic’s annualized revenue reached $65 billion in July, per CNBC’s account of the investor update, up from the $47 billion the company disclosed in its own Series H announcement in May. Reporting relayed through Simon Willison’s summary of the Financial Times story adds that Anthropic expects a profitable third quarter and told investors it has 6,000 customers spending $100,000 or more annually. Note what that second figure implies: 6,000 accounts at the threshold would total $600 million, under 1% of a $65 billion run rate. The revenue lives far above the floor, in a small number of very large deployments — precisely the accounts with the leverage to refuse a 2× price step.

Follow the token, not the tier

Spend share flatters expensive models, because every dollar of Fable 5 buys half as much work as a dollar of Opus 5. Convert Anthropic’s July mix into implied volume and the flagship shrinks further.

Take each Anthropic model’s July spend share from Ramp and divide it by that model’s list output price from the pricing page — a crude but honest proxy for relative token volume. Anthropic’s models account for 63.4% of tracked spend in July, of which Fable 5 is 12.6%. On the volume proxy, Fable 5 falls to roughly 5.3% of Anthropic’s tokens, while Opus 4.8 rises from 44.2% of the dollars to about 37.3% of the tokens and the Sonnet line, at $10 to $15 per million output tokens, climbs from 18.7% of dollars to over 30% of volume. The premium tier is not merely unpopular; it is a small slice of actual work wearing a large slice of the bill.

The same method yields the number worth quoting at a renewal. Weighted by July’s spend mix, Anthropic’s blended list price works out to $21.12 per million output tokens — against $19.18 in June and $19.83 in March. Anthropic’s average realized price per unit of work went up about 10% in one month, not because it raised prices but because Fable 5 pulled the mix upward at the same time cheaper tiers pulled volume the other way. Anyone modeling next year’s spend from a flat “average price of Claude” is modeling a number that moves whenever a launch reshuffles the mix.

Against that, the competitive picture in the same dataset is less comfortable. Summing model-level shares by provider, Anthropic held 63.4% of tracked US business model spend in July versus OpenAI’s 36.0% — a slide from 71.4% to 28.1% a month earlier, driven almost entirely by GPT-5.6 Sol arriving at 14.88% of all spend in its first full month. That is the fastest single-model share gain anywhere in the July panel, and it landed while Anthropic’s own new flagship was struggling to clear 8%.

The price war behind those movements is explicit. As Ars Technica’s account of the FT reporting sets out, OpenAI cut GPT-5.6 Luna from $1 to $0.20 per million input tokens and from $6 to $1.20 on output, while Anthropic called off a planned September increase for Sonnet 5. Both moves attack the middle of the range. Neither touches the top. Mantas Lukauskas, AI tech lead at the hosting company Hostinger, gave the same piece the cleanest reading of the pattern: “The US labs have cut the middle and are defending the top.”

The defended top has a shape worth naming. OpenAI’s published API pricing puts GPT-5.6 Sol at $4 and $20 per million tokens on short context, doubling to $8 and $30 on long context, with the note that Sol’s promotional pricing runs at least through November 21, 2026. Anthropic’s Fable 5 sits above both at $10 and $50 flat. So the premium band spans roughly $20 to $50 per million output tokens while the volume band — Sonnet, Haiku, Luna — runs $1.20 to $15. Buyers are not choosing between two prices; they are choosing between two orders of magnitude, and July’s mix says the mass sits in the lower one.

Where this reading could be wrong

The most serious objection is that per-token price is a bad proxy for the cost of finishing a job, and there is hard evidence for it. Databricks benchmarked models and harnesses against its own multi-million-line codebase and found in its coding-agent benchmark results that Sonnet 5, roughly 1.7× cheaper per token than Opus 4.8, cost $2.09 per task against Opus’s $1.94 while scoring six points lower, because it consumed 1.9× more tokens getting there. If that inversion holds at the top of the range, Fable 5’s 2× token premium could be cheaper per completed task, and the spend data would be measuring buyers’ misunderstanding rather than the models’ economics.

Anthropic’s own materials argue the opposite direction, which weakens the defence. The Opus 5 announcement claims the cheaper model matches Fable 5’s peak score within 0.5% on one coding benchmark at half the cost per task, and beats it on a computer-use benchmark at just over a third of the cost, citing evaluations including Frontier-Bench and the GDPval-AA leaderboard run by Artificial Analysis. When the vendor’s own cost-per-task charts favour the cheaper tier, buyers moving away from the expensive one look well informed, not confused.

Three other caveats deserve stating plainly. Ramp’s panel covers corporate-card and bill-pay transactions at 70,000-plus US businesses, which skews toward startups and mid-market firms and away from the negotiated enterprise contracts where the largest Claude deployments live — the 6,000 six-figure accounts may buy an entirely different mix. Fable 5 has had two months in the data against Opus 4.8’s four, and a slow-burn flagship is a real pattern. And Anthropic’s revenue and customer counts here come from unnamed sources in press reporting rather than a filing, with an IPO process that will eventually replace all of it with audited numbers.

The evidence that would flip the verdict is specific: a September or October reading showing Fable 5 above roughly 15% of spend, or a credible cost-per-task benchmark where the premium tier finishes work for less money than Opus 5. Absent either, the mix says what it says.

What to do before your next renewal

The operator question is not which model is best. It is which tier your workloads actually require, and whether your contract prices that honestly. A vendor defending its top tier while cutting the middle is a vendor that will happily let a default setting do its selling.

Audit your tier mix against your task mix this week. Pull thirty days of spend by model and compare it to task difficulty. The Databricks work found roughly a quarter of engineering requests were low-complexity and about 60% medium; if your premium-tier share exceeds your genuinely hard-task share, the gap is pure margin transfer.
Benchmark cost per completed task, never cost per token. Run the same fifty tasks from your own backlog through both tiers and compare total spend to pass rate. This is the only number that settles the Fable-versus-Opus question for your codebase, and it costs a few hundred dollars to produce.
Renegotiate on mix, not on rate. The blended $21.12 figure moves with every launch. Ask for committed-spend pricing that holds across the family rather than per-model discounts that evaporate the moment the vendor renames its flagship.
Watch for the default flip. Opus 5 became the default on Claude Max at launch; defaults are where mix changes get made without a purchase order. Pin model versions in production and review the pin monthly.
Price the second supplier. GPT-5.6 Sol took 14.88% of spend in one month, and the archive’s analysis of Chinese open-weight pricing pressure on US enterprise buyers shows the floor keeps falling. A live fallback is the cheapest negotiating instrument available.

Two other stories in today’s edition tell the same story from the funding side: SoftBank is selling ¥1 trillion of retail bonds to meet its OpenAI commitments, and Alibaba priced a $10.2 billion placement to fund AI infrastructure. Capital is being raised at record scale on the assumption that model revenue compounds at premium prices. The July mix is the first clean evidence that buyers are compounding it at the cheaper tier instead — a tension this paper first flagged in Anthropic’s IPO math of a $65B run rate against a $42B loss, and one that also shapes how much a discounted token is really worth once you verify what served it.

The flagship nobody buys

Follow the token, not the tier

Where this reading could be wrong

What to do before your next renewal

Sources