Anthropic's Opus 4.8 dominates model spend as new Opus 5 holds 3.5%

Key facts

• Anthropic's annualized revenue reportedly reached as much as $65bn in July 2026, up from $47bn in May, according to unnamed sources cited by the Financial Times.
• Anthropic told investors it has 6,000 customers spending $100,000 or more a year, and expects the third quarter to be profitable using the same method it used to call the second quarter profitable.
• OpenAI's annualized revenue reportedly climbed 35% quarter to date to more than $40bn, which the FT partly credits to July's launch of GPT 5.6 after a slow start to the year.
• Ramp's AI index, drawn from July billing data across 70,000 companies, shows Opus 4.8 taking 28.0% of Anthropic model spend, more than three times any other single version.
• Opus 5 held just 3.5% of July spend, which Willison calls reasonable since it was available for only about a week of that month after its 24 July release, while Fable 5's comparatively low 8.0% share is one he attributes to cost.

Why it matters

Two of the best-funded AI labs are reporting fast revenue growth this year. Anthropic's annualized figure reportedly reached as much as $65bn in July, up from $47bn just two months earlier, and OpenAI's climbed 35% quarter to date to more than $40bn. Anthropic is also telling investors it expects to stay profitable into the third quarter. But the Ramp billing data complicates the growth story: it shows that within Anthropic's own lineup, spend is heavily concentrated on an older model, Opus 4.8, rather than on the newest release. That gap matters because headline revenue growth does not by itself say which specific models businesses are actually choosing to pay for, and the answer here is not simply the newest one.

Who it affects

Anthropic and OpenAI are affected most directly: their reported revenue trajectories and profitability claims shape how investors value them ahead of any future fundraising. Enterprise customers and finance teams who pay for AI models through corporate cards matter too, since their aggregate billing habits are what the Ramp index actually measures. The numbers also matter inside Anthropic itself, for its product and pricing teams: a newly launched flagship pulling just 3.5% of July's spend despite being on the market for only about a week of that month, and Fable 5 landing well behind an older Opus version, are both signals about what is driving adoption inside its own lineup.

How to use it

There is no product to buy here, but the numbers work as a reality check for anyone budgeting for AI spend: on Ramp's evidence, a large share of what companies actually pay for still goes to an established, previous generation model, Opus 4.8, rather than to Anthropic's newest release. Teams weighing whether to move workloads to a brand new model on day one, instead of waiting, can read this as a sign that many peers are not moving immediately either, for reasons ranging from price to simple inertia.

How solid is it

Every revenue, customer and profitability figure here traces back to unnamed sources, described only as people with knowledge of the matter. They reach this retelling secondhand: from those sources to the Financial Times, then to Willison's post. None of it is confirmed by an Anthropic or OpenAI filing or a quoted company statement. The Ramp figures rest on different evidence: actual card billing data across 70,000 companies. That is a real usage signal, not a survey, but it is still a proxy, since it only sees spend that runs through Ramp's own corporate cards. The ten listed model shares add up to about 63%. Neither Willison nor the underlying report explains what the remaining spend covers.

Risks and caveats

No individual at Anthropic, OpenAI, Ramp or the FT is named in what Willison relays, so none of the claims can be checked against a specific, identifiable source. OpenAI's 35% revenue jump is not tied to a stated baseline, so it is unclear whether it is measured against the prior quarter or the same period a year earlier. Anthropic's third quarter profitability is described as an expectation, not a result, and the method it says it used to call the second quarter profitable is left unexplained, so the claim cannot be independently checked. The Ramp breakdown gives percentage shares only, with no dollar figures, so it cannot say whether the underlying market is growing or shrinking, only how one company's spend currently splits across its own model lineup.