Anthropic's revenue has outpaced companies that took decades to build, and the numbers deserve scrutiny
Anthropic is being discussed with the kind of revenue figures that make long-established companies look slow. But how that revenue is counted matters as much as how large it has grown.
Anthropic’s revenue figures are drawing comparisons that would have seemed implausible even 18 months ago. Michael Batnick, Director of Research at Ritholtz Wealth Management, puts the company’s annualized revenue run rate ahead of Pfizer, Cisco, and Intel, and describes it as $10 billion ahead of Uber and $15 billion ahead of Coca-Cola. Those are companies that spent decades accumulating the customers, infrastructure, and brand recognition that produce that kind of revenue. Anthropic has existed for only a few years.
The paired figure from Nathaniel Whittemore, Founder and Chief Executive of Superintelligent, widens the lens further. Combined with OpenAI, the two companies are running at $100 billion in annualized revenue, up from low single-digit billions at the same point last year. That rate of expansion does not have many precedents in markets of any kind, and the framing keeps appearing across financial and technology discussions.
Before those figures are taken at face value, the methodology behind them warrants attention. Whittemore notes that Anthropic calculates its annualized run rate by taking the past four weeks of application programming interface revenue and extrapolating that to a full year, rather than deriving it from recurring contracts. That is a meaningful distinction. A run rate built on four weeks of API activity reflects momentum, not locked-in demand, and a deceleration in any given month would revise the implied annual figure substantially. The headline number is real, but its construction makes it a forward projection as much as a present measurement.
The revenue conversation is not the only financial angle worth tracking. Robin Wigglesworth, Editor of FT Alphaville at the Financial Times, flags a subtler dynamic in how large technology companies report earnings. A meaningful share of what Microsoft, Google, and Amazon currently report, he argues, derives from marking up the value of their private stakes in Anthropic, OpenAI, and SpaceX rather than from operating performance. If that reading is accurate, then Anthropic’s valuation growth is feeding back into the reported earnings of its own largest backers, creating a circular quality that conventional earnings analysis was not built to handle.
Between these two companies, you're talking about a hundred billion in revenue run rate, up from low singledigit billions at this time last year. Nathaniel Whittemore
On the technical side, security researchers are examining structural properties of Anthropic’s models that have nothing to do with revenue. Ilia Shumailov, an AI and security researcher at Meta, describes a shared vulnerability across Anthropic, OpenAI, and Google: the reasoning traces of larger models in a given family can be replayed into smaller models in the same family, enabling what he calls trace-extraction attacks. This is not a flaw unique to Anthropic, but its presence across all three major providers suggests it is an architectural pattern rather than an isolated oversight.
A separate set of findings concerns how different models respond to adversarial prompts designed to induce specific beliefs. Nathan Labenz, an AI analyst and entrepreneur at Waymark, describes Anthropic research on what the company calls a mind-virus payload, specifically an AI supremacy framing. In those tests, certain models showed susceptibility to the payload while others did not. The result is preliminary and the framing is Anthropic’s own, but it points to an emerging research area around model belief manipulation that goes beyond standard safety benchmarks.
There is also a capability gap between what Anthropic publishes and what it runs internally. Labenz notes that the company’s internal Model 2 scores 8.8 percentage points higher on CobBench than the publicly released Methos Preview. A gap of that size, if the benchmark is meaningful, suggests that the publicly available models are not the ceiling of what Anthropic has built.
Taken together, the picture is one of a company whose scale is now large enough to distort adjacent financial reporting, whose internal capabilities may exceed its public releases by a measurable margin, and whose models share security-relevant architectural properties with competitors. The revenue figures are the loudest part of that picture, but they are not the only part worth watching.