One of These Companies Makes Money. It’s Also the One Raising Your Prices.

Bloomberg reported on August 15 that Anthropic posted preliminary Q2 2026 revenue above $11.5 billion, up from $787 million in the same quarter a year earlier and $4.73 billion in Q1. It also recorded positive adjusted operating income, its first. Meanwhile OpenAI filed a confidential S-1 in May targeting a September listing at up to $1 trillion, having lost roughly $1.22 for every dollar it earned in Q1 2026, with profitability not expected until around 2030. Both are heading for public markets within months of each other. The practical consequence for anyone paying for these tools is that pricing stops being a growth lever and starts being a reported line item. Sonnet 5 introductory pricing ends August 31. Best for understanding why your subscription costs are moving. Not investment advice.

Back in May, Anthropic told investors it expected $10.9 billion in Q2 revenue.

The actual number came in above $11.5 billion.

A year earlier, that same quarter produced $787 million. So revenue grew more than fourteenfold in twelve months, according to documents Bloomberg reviewed and CNBC reported on August 15.

Fifteen days from now, the price of Sonnet 5 goes up.

Those two facts are the same story. Almost nobody is telling it that way.

The Numbers, With the Caveats Attached

Being careful here, since two different sets of figures are circulating and people keep mixing them up.

Two things in that table deserve flagging rather than glossing.

The Anthropic figures are preliminary and could still change. Bloomberg said so explicitly.

And “positive adjusted operating income” is not the same as profit. Adjusted means somebody chose which costs to exclude. The May projections put operating profit at $559 million while including model training costs and excluding stock-based compensation, which is a specific set of choices. Until an S-1 becomes public, nobody outside these companies knows exactly what got adjusted out.

The Caveat That Actually Matters

Certainly there is one criticism of Anthropic’s profitability worth taking seriously. It comes from the reporting rather than from cynics.

Notably, Anthropic pays SpaceX roughly $1.25 billion a month for compute, disclosed in SpaceX’s own S-1. That contract carries a ramp-up discount, meaning Anthropic pays a reduced rate during the early months.

Those early months are Q2 2026. The exact quarter in which profitability appeared.

Tech journalist Ed Zitron published a critique titled “Anthropic’s Profitability Swindle” arguing the profit figure is a direct product of that reduced rate. Anthropic itself has indicated the profitability will not be sustained.

So the honest version is that Anthropic reached profitability in a quarter where its largest cost was temporarily discounted, then expects to lose it again when the full rate applies. That is meaningfully different from a company whose economics have permanently flipped.

Still, it is not nothing. Compute spending per revenue dollar fell from 71 cents to 56 cents in a single quarter. Some of that is the discount. Some of it is revenue growing faster than costs, which is the thing that has to happen eventually for any of this to work.

What OpenAI’s Numbers Look Like

A different shape entirely, though.

Currently, roughly $2 billion in monthly revenue, hitting about a $25 billion annualised run rate. Nearly $6 billion in Q1 alone. More than 230 million weekly ChatGPT users. Those are enormous numbers by any measure.

Against that, a negative 122% non-GAAP operating margin in Q1 2026. Every dollar of revenue came with $1.22 of loss attached, working out to roughly $6.95 billion in quarterly losses on training and inference. Projected operating losses for 2026 sit around $14 billion. The company does not expect positive cash flow until roughly 2030.

The reason is compute. Those commitments are extraordinary. Around $50 billion on infrastructure in 2026 alone. Amazon Web Services at $38 billion over seven years. Oracle at $60 billion a year for five years starting 2027, totalling $300 billion. AMD supplying 6 gigawatts of Instinct GPUs.

Obviously none of that is irrational if the revenue arrives. Sam Altman has publicly targeted $100 billion in annual revenue by 2027. The whole valuation depends on believing that number.

Why Both of Them Filed Within Ten Days

Firstly, OpenAI filed confidentially on May 22. Anthropic filed on June 1.

That is not a coincidence, it is a race. Whoever lists first sets the pricing benchmark for everyone behind them. A strong OpenAI debut at $1 trillion builds the runway for Anthropic. A flat one at $852 billion compresses every AI listing that follows.

Meanwhile SpaceX filed too, targeting somewhere between $1.75 and $2 trillion after absorbing xAI. Three frontier companies hitting public markets in the same quarter has no precedent.

The useful part for everyone outside these companies is the disclosure. A confidential S-1 stays sealed until roughly fifteen days before the roadshow. When it opens, we get the first audited look at what frontier AI actually costs to run. Real margins, real contract obligations, real compute economics, all of it previously hidden behind private company walls.

That document will settle arguments that have run on speculation for three years.

What This Means for Your Subscription

So here is the part affecting you rather than the analysts.

Generally, private companies price for growth. Losing money to acquire users is a strategy investors reward, because the story is about market share and the losses are somebody else’s problem until the next round.

Public companies price for margins. Every quarter, a number gets reported. The people who own the stock then have opinions about it. Pricing stops being a growth lever and becomes a line item somebody has to defend on an earnings call.

Both of these companies are weeks away from that transition.

Now look at what has already happened in the last month. Sonnet 5 introductory pricing at $2 and $10 per million tokens ends August 31, moving to $3 and $15. DeepSeek scrapped flat API pricing yesterday for peak and off-peak rates that run higher even at the discount, plus raised V4 Flash by 93%. Grok 4.6 launched with a clause that reprices your entire request at double rate if it exceeds 200K tokens.

Meanwhile OpenAI cut Luna by 80% and made ChatGPT’s free tier unlimited, which looks like the opposite move until you notice it is buying users ahead of a listing.

The pattern across all of it: cheap acquisition at the top of the funnel, rising prices where the usage actually is. Our Claude Pro review covers what the consumer tier gets you today. Today is the operative word.

The Thing Nobody Says Out Loud

Anthropic’s revenue engine is not the chatbot, incidentally.

The reporting is consistent on this: enterprise adoption of Claude Code drives the growth, with over a thousand enterprise accounts spending more than a million dollars each. High-margin recurring API revenue, generated by developers running agentic coding workloads at scale.

Which reframes the last three weeks of coverage on this site, actually. When DeepSeek gave away an entire agent runtime for free then raised model prices four days later, that was an attack on exactly this revenue line. When Prime Intellect published a self-improving harness under MIT, same layer.

The agent tooling war and the profitability numbers are the same war. Whoever owns the runtime that enterprises run their coding workloads through owns the revenue that just made one of these companies profitable.

That is why a frontier lab gives away a $0 product in the same week it raises prices on the model underneath.

Anthropic’s Other Move This Week

One more thing landed in the same news cycle and it fits the pattern.

Anthropic is reportedly acquiring a startup for $6 billion. The target has not been widely confirmed, so treat that figure as reported rather than settled until somebody names the company.

A $6 billion acquisition from a company that has just posted its first profitable quarter, weeks before a public listing, is a specific kind of signal. Acquisitions before an IPO get scrutinised heavily in the prospectus, because they change the story a company is telling about how it grows.

Spending six billion dollars right after establishing you can make money suggests confidence that the money will keep arriving. It also suggests something the company wants badly enough to buy rather than build, which is usually talent, distribution, or a capability that would take too long internally.

Worth watching what gets named, since the answer says more about where Anthropic thinks the next revenue comes from than any earnings figure will.

Why the Growth Rate Is the Real Story

Set the profit argument aside momentarily, because the revenue curve is harder to dispute than the accounting.

Anthropic did $787 million in Q2 2025. It did over $11.5 billion in Q2 2026. That is more than fourteen times in twelve months. None of it depends on how you define adjusted anything.

Analysts have pointed out this growth rate exceeds Zoom at its pandemic peak, plus Google and Facebook in the quarters before their own listings. Those are the historical comparisons people reach for when something moves unusually fast. This moves faster.

Previously, less than a year ago, the company was telling investors not to expect full-year profitability before 2028. That guidance is now four years out of date, based on data shared with the same investors.

The driver is not consumer subscriptions. Reporting consistently identifies enterprise adoption of Claude Code, with over a thousand accounts spending seven figures each. That is recurring high-margin API revenue from developers running agentic workloads, which is a completely different business from a chatbot with a monthly fee.

Worth holding both thoughts. The profit may be a contract timing artifact. The revenue trajectory is not. That is the number deciding whether Q3 recovers.

What Happens When These Become Public Documents

Ultimately the most useful thing about both filings has nothing to do with share prices.

Frontier AI economics have run on speculation for three years. Nobody outside these companies knows what inference actually costs at scale, what margins look like on enterprise contracts, or how much of the revenue is real versus committed credits from investors who are also suppliers.

An S-1 forces all of that into daylight. Audited financials, contract obligations, customer concentration, related-party transactions. Every argument that currently ends in “well, we don’t know” gets an answer.

Some of those answers will be uncomfortable, certainly. Compute commitments that read like sovereign debt. Revenue concentrated in a handful of enterprise accounts. Circular arrangements where an investor is also a supplier is a structure that looks very different in a prospectus than in a press release.

For anyone building on these APIs, that transparency is worth more than the stock. A public company has to disclose material changes to its business, which includes the economics behind the pricing you depend on.

Right now you find out your model got more expensive when a blog post appears. After the listing, the reasoning behind it lands in a quarterly filing.

What to Watch

Three checkable things over the next two months.

Firstly, whether the public S-1 confirms the adjusted numbers. Preliminary figures with an adjustment somebody chose are not audited financials. The full prospectus is where the compute obligations, the SpaceX contract terms, plus the actual definition of adjusted operating income all become public.

Secondly, whether Q3 holds. Anthropic has signalled the profitability will not be sustained once the SpaceX ramp-up discount expires. If Q3 stays positive anyway, the economics genuinely flipped. If it does not, May’s milestone was a contract timing artifact.

Thirdly, whether the September window opens. SEC review of confidential filings typically runs 60 to 90 days. Market conditions also have to cooperate. A delayed OpenAI listing pushes everything behind it.

What This Looks Like From Where You Sit

Strip out the valuations plus the filings, because most of that is somebody else’s problem.

The practical situation is that four companies supply almost all the model capacity anyone uses, two of them are about to become publicly traded, plus all four repriced something in the last four weeks.

That is not a coincidence of timing. It is what happens when the industry moves from a phase where investors fund the losses to a phase where customers do.

Nothing about this is unusual, historically. Cloud computing went through it. Streaming went through it. The pattern is always the same: land grab pricing while somebody else pays, then a correction once the growth story needs a margin attached to it.

The difference here is simply speed. AWS took roughly a decade to get from launch to price discipline. This took about three years.

Which means the practical advice is unromantic. Assume today’s price is a floor rather than a ceiling. Anything you build that depends on a specific cost per million tokens needs a plan for that number doubling, because it has now doubled for several vendors inside one month.

Where that argument leads is toward the open-weight side, which is exactly why the agent tooling war matters. If the runtime is swappable and the model is a configuration line, a price rise becomes an inconvenience rather than a crisis. If your workflow is welded to one vendor’s product, it becomes a budget conversation.

That is the real reason to care about any of this. It has nothing to do with whether anyone should buy the stock.

The Part Worth Keeping

For three years now, the standard argument against AI companies was that nobody had shown the business could work. Enormous revenue, larger costs, profitability always a few years out.

One of them just posted a profitable quarter, in a quarter where its biggest bill was temporarily discounted, while telling everyone it will not last.

The other is heading to the largest technology listing in history while losing $1.22 on every dollar it earns.

Both are about to have to explain themselves in public, quarterly, to people who own the stock. Whatever the disclosures say about compute margins and contract obligations, the part that reaches you is simpler.

The price of the tools goes up. It was always going to.

The only real question is whether you built something that can move when it does.

Charts and Blocks

Two companies, two shapes

What changed in your bill this month

FAQ