Anthropic Passes OpenAI on Revenue Ahead of IPO

For the first time since the generative AI market became an investable category, the company with the most revenue is not OpenAI.

Anthropic booked more than $11.5 billion in preliminary revenue in the second quarter, against $6.7 billion at OpenAI over the same three months. The gap is not a rounding difference or a definitional quirk at the margins. It is roughly 1.7 times, and it opened up in a single quarter: Anthropic's first-quarter revenue was $4.73 billion, so the sequential increase was more than 140%, while OpenAI grew 18% off a $5.7 billion base.

Bloomberg first reported the figures, which are preliminary and drawn from documents circulated as both companies prepare to go public. Both have filed confidentially with the Securities and Exchange Commission. Anthropic is expected to price first, in September or October, with Morgan Stanley, Goldman Sachs and JPMorgan working on the offering.

The run rate is the number investors are actually trading on

Quarterly revenue is the audited-style figure. The number circulating in investor conversations is the annualized run rate, and Anthropic's passed $65 billion at the end of July, up from $47 billion in May and roughly $9 billion at the end of 2025. OpenAI's most recently disclosed run rate, shared internally by co-founder Greg Brockman in mid-August, is $40 billion, doubled from $20 billion at the close of last year.

Anthropic added roughly $56 billion of annualized run rate in seven months, while OpenAI added $20 billion over the same period. That is the single comparison that explains the repricing of both companies in private markets, and it is worth stating plainly because run-rate figures are otherwise easy to wave at. Anthropic's investors expect the company to close 2026 somewhere between $100 billion and $120 billion, according to the Financial Times.

A caveat that belongs next to every one of these numbers: the two companies do not necessarily calculate run rate the same way. A run rate is an extrapolation, typically from a recent month or a recent quarter, and the choice of window is a management decision, not an accounting standard. Neither figure has been through an audit. Investors buying either IPO will get a prospectus with real financial statements, and the relationship between those statements and the run rates being quoted today is one of the more interesting open questions in this cycle.

Profitability is where the divergence gets uncomfortable

Revenue leadership can be bought. What makes this quarter difficult for OpenAI is that it was not.

Anthropic recorded an adjusted operating profit of about $559 million in the second quarter. OpenAI's operating loss widened to $12.3 billion, up from $9.3 billion in the first quarter, meaning losses grew faster than revenue did. For context on the scale of that trajectory, OpenAI posted a net loss of $38.5 billion in 2025 on $13.07 billion of revenue.

The adjusted profit figure deserves the same scepticism as the run rate. Anthropic excludes stock-based compensation from it, which for a company of this size and vintage is not a small exclusion, and the company has not detailed the rest of its methodology. Adjusted operating profit at a private AI lab is a management-defined number. It is still directionally meaningful, because the direction is opposite to the competitor's.

The mechanism most often cited for the difference is inference efficiency. "Anthropic was much more token efficient than OpenAI but OAI has closed some of the gap," Gavin Baker of Atreides Management told Axios. Harrison Rolfes of Pitchbook framed the same point from the buyer's side: a premium-priced model that returns a correct answer the first time can cost less per completed task than a cheap model the customer has to re-run, or pay a human to check.

That framing matters more than it sounds. It moves the competitive question from price per million tokens, where the market has been in a race to the bottom for two years, to cost per completed task, where correctness is the dominant term. The quarter's real lesson is that in enterprise AI the revenue leader is now also the one losing less money to earn it, which is not how a land-grab market is supposed to work.

Coding is the wedge, and enterprises are the buyer

The commercial engine behind Anthropic's quarter is not a consumer chatbot. It is developer tooling and enterprise API consumption, with Claude Code the most visible line item. Coding workloads have three properties that make them unusually good revenue: they are high-volume, the output is verifiable, and the buyer is a company with a budget rather than an individual with a subscription.

OpenAI has not missed this. It is on track to generate more than half its revenue from enterprise customers by year end, and it has bundled ChatGPT, Codex and a browser into a single product it says is gaining users quickly, though it has published no supporting data. What it has also had is turnover. Chief Revenue Officer Denise Dresser departed after less than a year, following former COO Brad Lightcap and Fidji Simo, who had been widely read as a possible successor to Sam Altman. Brockman has stepped up his day-to-day involvement in product and commercial operations.

Executive churn at the exact moment a growth story stalls is the pattern public-market investors are trained to notice, and OpenAI will be filing into that scrutiny with a $852 billion valuation to defend. Anthropic raised its Series H-1 at a $965 billion valuation in May.

What to watch between now and pricing

Three things will determine whether this quarter reads as an inflection or as noise.

First, whether Anthropic's growth is durable or front-loaded by a small number of very large enterprise contracts. A 140% sequential jump usually contains at least one deal that will not repeat.

Second, whether the efficiency gap survives contact with scale. Both companies have signed deals with inference providers and both have chip ambitions; if OpenAI closes the token-efficiency gap, the profitability divergence narrows without either company changing its product.

Third, gross margin disclosure. The prospectuses will, for the first time, show what these models actually cost to serve. Every valuation currently in the market is an assumption about that number.

For now the ordering has changed, and it changed on operating discipline rather than on price. That is the part of this quarter that is hardest for a competitor to answer.

Cover image: New York Stock Exchange, Broad Street (CC0, via Wikimedia Commons).