On paper, the AI industry's biggest upset is settled: Anthropic, the Claude maker long cast as OpenAI's smaller, safety-obsessed rival, now says its business is throwing off revenue at nearly twice OpenAI's pace. The company disclosed in late May that its run-rate revenue crossed $47 billion, up from roughly $9 billion at the end of 2025. OpenAI's most recently reported figure is about $25 billion. By that yardstick, Anthropic didn't just catch OpenAI — it lapped it.
But the yardstick is the whole story. Both numbers are run-rates, not audited annual revenue, and the tidy "Anthropic overtakes OpenAI" line papers over an accounting fight, a measurement quirk, and a comparison that mixes together things that don't belong in the same column. Here is what can and can't be confirmed as of Friday.
What the numbers actually say
Start with what is solid. The $47 billion figure comes straight from Anthropic — specifically its May 28 announcement of a $65 billion Series H round at a roughly $965 billion valuation. "Since our Series G in February, adoption has continued to grow across global enterprise customers, and our run-rate revenue crossed $47 billion earlier this month," the company wrote. It is a self-reported number, but one attached to a fundraise, which raises the stakes on accuracy. As independent analyst Simon Willison noted, "lying to investors who just put in $65 billion would be securities fraud," and the real figures will surface in Anthropic's IPO filing regardless.
OpenAI's ~$25 billion run-rate was reported by The Information, which said the company topped that mark "as of last month," a 17% increase over the $21.4 billion it was generating at the end of 2025. That, too, is credible — but notice the timing. Both data points are from spring 2026. No fresh revenue figure landed on July 17. The nearest recent news is that Anthropic has begun lining up banker and investor meetings ahead of a possible IPO (CNBC, July 15). The "overtake" is real, but it is weeks old, not a Friday headline.
Now the part the viral framing gets wrong. The claim circulating this week pits Anthropic's $47 billion against OpenAI's "$25–33 billion." The $33 billion figure is not revenue. It traces to a projected GAAP loss estimate from the forecasting firm FutureSearch — a very different thing. OpenAI's credibly reported 2026 loss is closer to $14 billion (The Information). Comparing one company's revenue run-rate to a garbled blend of another's run-rate and its loss forecast is not a scoreboard; it's a category error.
Why "run-rate" deserves scare quotes
A run-rate is an annualized projection: take recent revenue and extrapolate. The trouble is in the details. Reuters Breakingviews columnist Karen Kwok reported the mechanics of Anthropic's version, citing a person familiar with the matter: the company takes "the last 28 days of sales from customers charged on a consumption basis and multiply it by 13," then adds monthly subscription revenue multiplied by 12. That multiply-by-13 quirk alone inflates the consumption side above a straight annualization.
Critics have been blunter. "Annualized recurring revenue is one of the most regularly-abused statistics in the startup world," writes commentator Ed Zitron. "If you use ARR, you're essentially taking one month and treating it as representative of the entire calendar year, when it isn't" — no adjustment for churn or seasonality.
OpenAI, unsurprisingly, agrees when the target is Anthropic. In an April memo to shareholders, OpenAI Chief Revenue Officer Denise Dresser accused Anthropic of an "inflated" run-rate and of using accounting methods to make its "revenue look bigger than it is," alleging Anthropic grosses up revenue shared with cloud partners Amazon and Google while OpenAI reports its Microsoft share net. Anthropic has not conceded the point. The dispute is itself a signal: these are private companies disclosing selectively, and each frames the math to flatter itself.
The real divergence: enterprise versus consumer
Strip away the accounting theatrics and a genuine difference remains — one that explains how Anthropic pulled ahead on revenue while trailing badly on users. OpenAI is a consumer juggernaut, with ChatGPT past a billion monthly users; the bulk of its revenue comes from consumer subscriptions. Anthropic barely competes for that audience. Its business is overwhelmingly enterprise and API — companies wiring Claude into coding tools, legal work, and financial services — a mix analysts peg at roughly 80% enterprise, though that precise split is not company-confirmed.
Enterprise revenue is higher-margin and stickier, and it scales with usage rather than headcount. That is why Anthropic can out-earn a rival with a fraction of the consumer reach. Co-founder Daniela Amodei frames the road ahead around capital intensity, not consumer growth: "It's a really big upfront cost to train the models and to serve inference on them," she said at Bloomberg Tech, arguing that frontier labs "are just going to need access to capital, and I think the public market is very well suited to that." Axios's Jim VandeHei, writing when Anthropic was merely at $30 billion, said he could not find "any company — in any industry, in any era — that has scaled organic revenue this quickly at this level."
What to watch next
The honest verdict: on disclosed run-rate, Anthropic has indeed overtaken OpenAI, and the enterprise-first strategy behind it is real and material. But "annualized revenue" is doing heavy lifting — these are self-reported, differently calculated projections, not recognized full-year results, and at least one widely repeated version of the comparison smuggles in a loss forecast as if it were revenue.
The tiebreaker is coming. Anthropic has filed confidentially for an IPO, and its S-1 will replace run-rate marketing with audited numbers under scrutiny. If OpenAI follows toward its own rumored 2027 listing, the two companies' revenue, churn, and margins will finally sit in the same, comparable column. Until then, treat the scoreboard as provisional — and read the footnotes.
"It's a really big upfront cost to train the models and to serve inference on them."— Daniela Amodei, Co-founder and President, Anthropic