Anthropic Passed OpenAI on Revenue. Coding Agents Are Why.

On May 28, 2026, Anthropic announced a $65 billion Series H at a $965 billion post-money valuation and mentioned, almost in passing, that its run-rate revenue (the last period's revenue multiplied out to a full year, not actual trailing twelve-month revenue) had crossed $47 billion earlier that month.^[1] OpenAI, over the same stretch, was reported at roughly $25 billion annualized as of March, with an internal full-year target somewhere in the low thirties.^[5]

Sit with the shape of that. ChatGPT has something on the order of 900 million weekly users. Claude has a rounding error of that by comparison. And Anthropic is booking close to twice the revenue. The company that won the consumer war is losing the revenue war to the company that never really fought it.

This is not a story about model quality. It's a story about which customer you decided to serve, and what that customer's bill looks like at the end of the month. The way I think about it: OpenAI built a subscription business and Anthropic built a metered utility. Those are not the same company wearing different logos. They have different margins, different churn, different growth ceilings, and they get valued very differently once public-market investors get a vote.

The number that actually explains it is $2.5 billion

In February 2026, Anthropic said Claude Code's run-rate revenue had passed $2.5 billion, having more than doubled since January 1, with weekly active users doubling over the same six weeks.^[3] That was reported alongside a company-wide run rate of about $14 billion at the time, so Claude Code was pushing 18% of the whole business from a product that ships as a terminal command.

A terminal tool. Not an app, not a social feed, not a video model. A thing you install with a package manager and run in a directory. It went generally available in May 2025 and hit a billion-dollar run rate inside about six months. There is no enterprise software product in history with that curve, and I say that as someone who is instinctively suspicious of “fastest ever” claims.

The reason it works is boring and structural. A coding agent burns tokens (the chunks of text a model reads and writes, and the unit everything gets billed in) at a completely different order of magnitude than a chat message. A chatbot turn is maybe a few hundred to a couple thousand tokens. An agentic coding task, where the model reads files, runs tests, reads the failure, edits, re-runs, and re-sends the whole accumulated context on every single tool call, can run into the millions.^[9] Same model, same GPU, wildly different meter reading.

A $20 a month consumer subscription is a ceiling. It is the most that customer will ever pay you, no matter how much they love the product, and every extra query they run makes your gross margin worse. That is a genuinely strange business: your best users are your least profitable ones. A metered API seat is a floor. The more the developer uses it, the more they pay, and heavy usage means they've wired the thing into their build pipeline, which means they aren't leaving.

Why a CFO signs off on a coding agent and not much else

Here is the part I think most AI commentary misses. Coding agents are the killer app not because coding is intrinsically special, but because it is the one AI use case where the return on investment is legible to a finance person without a leap of faith.

A fully loaded senior engineer in the US costs a company somewhere north of $250,000 a year, call it $20,000 a month. If a coding agent costs $250 a month and makes that engineer even 5% more productive, the math clears trivially. You do not need a philosophy seminar about transformative AI. You need a calculator. Compare that to “knowledge workers can now summarize documents faster,” which is real, is valuable, and is completely impossible to defend in a budget review.

That legibility is the whole moat. It is why Anthropic could push into enterprise while OpenAI was optimizing a $8-a-month consumer tier to grow subscriber counts.^[12] Both are rational. Only one of them prints revenue per seat that looks like enterprise software.

The margin story underneath the revenue story

Revenue is the headline. Gross margin is the business. SemiAnalysis, which does the closest thing to real financial modeling on these private companies, has Anthropic's gross margin climbing from deeply negative in 2024 to roughly the 60% range in 2026, with the API-only slice running north of 80%, and projects Anthropic clearing $1 billion in quarterly operating profit in Q3 2026.^[7]Over the same window it has OpenAI's operating margin still in the neighborhood of negative 100%.^[7]

That gap is not about who trains better models. It's about who has to feed a free tier. Serving hundreds of millions of free ChatGPT users costs real money every single day and produces no revenue at all. It buys you distribution, brand, and a funnel. It also means that at any given revenue level, OpenAI is carrying a cost base Anthropic simply does not have. If both companies got to $100 billion in revenue tomorrow, they would not have the same gross profit, and it is not close.^[11]

Consumer AI is a share-of-attention business with search-engine economics but without search-engine ad revenue attached yet. That “yet” is doing enormous work in OpenAI's valuation.

What this does to the OpenAI IPO story

OpenAI filed a confidential S-1 (the registration statement a company submits to the SEC before going public) on June 8, 2026, reportedly targeting up to a $1 trillion valuation. Since then the New York Times has reported the company is leaning toward pushing the listing to 2027 rather than accepting a price below that trillion-dollar mark, with Sam Altman reportedly calling any cut a non-starter.^[6]

I understand the impulse. I also think it's the wrong read of what public markets will do here. The moment OpenAI files publicly, the S-1 shows everyone the revenue mix, the free-tier cost of goods, and the loss line. Reported forecasts have OpenAI losing something like $14 billion in 2026 and not reaching profitability until the back half of the decade.^[7] Meanwhile a direct competitor with fewer users, less mindshare, and a fraction of the marketing budget is booking more revenue at better margins and may be printing operating profit before the IPO even prices.^[7]

A $1 trillion valuation on ~$30 billion of revenue is roughly 33x sales. That is a number you can defend if you are the category-defining growth story with no credible challenger on the fundamentals. It is a much harder number to defend when the comparison sitting right next to you in the prospectus has better ones. Waiting until 2027 to protect the headline valuation only works if the gap closes. If it widens, delaying just means pricing into a market that has had twelve more months to internalize the margin story.

Where I think the bulls on this narrative go too far

Now the other side, because the “Anthropic won” take has gotten lazy fast.

• The $47 billion is self-reported and it's a run rate. Ed Zitron has written a pointed piece arguing Anthropic's profitability framing is doing a lot of selective accounting, and even if you think he's too harsh, he's directionally right that none of these numbers are audited.^[8] Nobody outside these companies has seen a real income statement.
• Revenue concentration is a real risk.If a huge share of Anthropic's revenue routes through coding, and a meaningful chunk of that through a handful of very large customers and API resellers, then one big enterprise deciding to cap spend, or one competitor shipping a good enough model at half the price, hits harder than a diversified consumer base would.
• Token prices fall.Metered revenue is wonderful right up until inference gets commoditized. Anthropic's margin improvement has come substantially from inference efficiency, which is the same force that eventually lets someone undercut you.^[7] The 80%+ API gross margin is a target for every competitor on earth.
• OpenAI is not standing still on coding. Codex has been growing fast on developer counts, and OpenAI has the distribution to bundle aggressively.^[5] Distribution is not nothing. Ask anyone who ever competed with Microsoft Office.

What I'd actually watch

Three things, and none of them is the next run-rate press release.

1. Net revenue retention on enterprise API accounts.This is the number that separates a real utility from a hype cycle. If enterprises are spending 130%+ of what they spent a year ago, the flywheel is real. If they're capping budgets like Microsoft's E+D group reportedly did, the growth is borrowed.^[9]
2. The first public S-1. Whichever of these two files first hands the entire industry a real cost structure for the first time. Free-tier COGS, compute commitments, customer concentration. That document will reprice more than one company.
3. Whether OpenAI monetizes attention.Ads, commerce, transactions. 900 million weekly users is an asset that Anthropic cannot replicate at any price, and it is currently generating close to nothing per free user. If OpenAI cracks that, this entire article ages badly, and I'd rather say so now than pretend otherwise.

Anthropic didn't out-engineer OpenAI into the revenue lead. It picked a different customer. Developers and enterprises pay by the token, use more when the product gets better, and can't easily rip the thing out once it's in the build pipeline. Consumers pay a flat twenty bucks and cost you more every time they hit enter. The same underlying technology, pointed at two different markets, produced two different companies. One of them is about to ask the public markets to value it like the winner. The other one already is.

If you build software for a living, the practical read is simpler and more immediate: your tooling budget is about to become a line item that finance actually looks at. That is what happens when a developer tool starts billing like a cloud provider.