Anthropic Went $787M to $11.5B in a Year. Then Came 'Adjusted'.

TL;DR

Anthropic booked preliminary revenue of more than $11.5 billion in Q2 2026, up from $787 million in the same quarter a year earlier and $4.73 billion in Q1 2026. It also posted positive adjusted operating income, which, as TNW put it, is not something the leading AI labs have been able to say. The figures come from documents shown to prospective investors ahead of a confidential IPO filing, first reported by Bloomberg on August 14. They are preliminary, they could be revised, and the word "adjusted" is carrying real weight.

The three numbers

Strip the IPO noise away and the quarter reduces to a single line that goes up very fast.

That is a 14-fold jump against the year-ago quarter and roughly 2.4x against Q1. Fortune and CNBC both carried the same set of figures from the same reporting, and every outlet attaches the same disclaimer: deliberations are ongoing, the numbers could move, and Anthropic declined to comment.

The run rate finally caught its own shadow

Four quarters at $11.5 billion annualizes to about $46 billion. Anthropic disclosed a $47 billion annualized run rate back in May. So the trailing number has now essentially caught up with the forward number that investors have spent all summer pricing off. The reported $2 trillion IPO chatter still rests on a 2026 revenue forecast rather than revenue in hand, but the base it extrapolates from is no longer hypothetical.

For context on the field: the same reporting puts OpenAI's run rate at over $40 billion. The gap between the two labs is now measured in single-digit billions, which is a sentence nobody was writing eighteen months ago.

The word doing the work is 'adjusted'

Here is where a builder should slow down. The investor materials circulating in May projected $559 million of adjusted operating profit on $10.9 billion of Q2 revenue, a margin of about 5.1%. That figure included model-training costs, which is the expensive and genuinely credible part, and excluded stock-based compensation.

Counting training but not equity is a specific choice. It is the restaurant that counts the food and the rent, then leaves the shares it handed the chefs off the tab. In a lab where nine-figure comp packages are the primary weapon in a talent war, that omission is not a rounding error.

And note what the August update actually confirms: the documents say adjusted operating income was positive. No public number has been attached to the higher $11.5 billion revenue figure. The sign is reported. The size is not.

Five cents on the dollar

Even taking the May projection at face value, the first operating profit a frontier lab has claimed is thin enough to see through.

Reuters Breakingviews made the obvious follow-on point: Anthropic may not stay profitable across the full year as compute spending climbs. At a 5% margin, one large new datacenter commitment landing in the wrong quarter flips the sign back.

Why this matters if you build on Claude

Three practical reads.

• Pricing stability, with limits. A lab that can cover its compute bill out of revenue has less reason to yank the tier you built on. That is consistent with Sonnet 5's intro price being made permanent. But a 5% margin also means there is no fat to give away, so do not budget for another 80% price cut.
• The token economics work at scale. Selling inference to enterprises, with Claude Code as the wedge, can pay for the training runs that produce the next model. That was an open question a year ago and it is now, at minimum, a demonstrated quarter.
• The S-1 is the real event. When the confidential filing goes public, it should carry the first audited gross margin ever disclosed by a frontier lab. Every capacity plan, pricing model, and buy-versus-self-host argument in this industry is currently built on estimates of that one number.

What is still unknown

Being precise about the gaps, because they are large:

• None of these numbers are audited. They are preliminary figures in investor materials, and the reporting says explicitly that they could be revised.
• No stock-based compensation figure has been reported, so the distance between "adjusted" and GAAP operating income is unmeasurable from the outside.
• No gross margin has been disclosed. Analyst models exist, but a model is not a disclosure.
• Anthropic is in a quiet period and declined to comment. The issuer cannot speak, so the documents speak for it, which is a strange way to run a $2 trillion price discovery.

Key Takeaways

• Anthropic's preliminary Q2 2026 revenue topped $11.5 billion, against $787 million in Q2 2025 and $4.73 billion in Q1 2026.
• The quarter carried positive adjusted operating income, reportedly a first among the leading labs, but no dollar figure has been attached to the higher revenue number.
• May investor materials projected $559 million of adjusted operating profit on $10.9 billion, about a 5.1% margin, including model-training costs and excluding stock-based compensation.
• Four quarters at this level annualize to roughly $46 billion, essentially matching the $47 billion run rate disclosed in May.
• Everything is preliminary and unaudited; the confidential S-1 going public is the moment real gross margin numbers arrive.

Sources: Bloomberg, Fortune, CNBC, TNW, R&D World, Anthropic