Anthropic filed its S-1 today. I pulled it down myself — 14 MB, 400 pages — because I had a hunch I wanted either confirmed or killed: that they are making a great deal of money on what we pay them, and that the price is coming down.
The headlines went for the net loss: $80.7 billion for the first half of 2026. It's a spectacular number and it's nearly meaningless. $72.4 billion of it is non-cash fair-value movement on convertible notes, derivatives and SAFEs (p. 63–64), and another $8.8 billion is fair value on the note purchase agreements (p. 150) — $81.1 billion together, against $1.1 billion of interest income (income statement, p. 147). Accounting noise from having raised $141 billion. The operating loss, the one that says something about the business, is $693 million.
The number I ended up staring at is somewhere else entirely.
They stopped printing the gross margin
Starting with fiscal 2025, Anthropic changed how its income statement is organised. It used to present expenses by function: cost of revenue, gross profit, R&D, sales and marketing, G&A. It now presents them by nature: compute and infrastructure, personnel, sales and partnerships, the Bartz settlement, other. Prior years were recast, so the comparison still holds.
The consequence is in their own words: "cost of revenue, gross profit (loss) … are no longer presented as separate line items in the consolidated statements of operations."
The gross margin still exists. It lives in the notes, under Supplemental Operating Expense Information on page F-39, in the one table I could find from which you can compute it:
| $M | 2023 | 2024 | 2025 | H1 2025 | H1 2026 |
|---|---|---|---|---|---|
| Revenue | 41 | 386 | 4,585 | 1,163 | 16,286 |
| Cost of revenue | 26 | 752 | 2,773 | 1,019 | 2,380 |
| Gross margin | 37% | −95% | 40% | 12% | 85% |
The margin row is mine; the filing never states it.
Read the 2024 column again. It cost them twice as much to serve a customer as the customer paid. Then read the last two columns: in twelve months they went from 12% to 85%.
The note under the table defines what cost of revenue is: "Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads." That's the cost of serving us. And footnote 3 on the same table puts the compute for free users under sales and marketing rather than cost of revenue. So the 85% is the margin on people who pay.
The entire loss is research and sales
Same table, first half of 2026, in millions:
- Cost of revenue: 2,380
- Research and development: 9,741
- Sales and marketing: 3,808
- General and administrative: 1,050
That's 16,979 against revenue of 16,286. The $693 million operating loss is the gap between a highly profitable serving business and a laboratory spending $9.7 billion to build the next model.
To be fair about what this is: a SaaS-shaped gross margin, not an all-in cost. Training sits in R&D, not in cost of revenue. Ask what Claude costs in total and the answer is still "more than they charge." Ask what it costs to serve one request from a paying customer and the answer is fifteen cents on the dollar.
The filing makes the same point in plainer language elsewhere: "our revenue grew 1,368% year over year for the three months ended June 30, 2026, compared to our compute and infrastructure expenses that grew 300% during the same period" (p. 136).
The subscription isn't what I thought it was
This is where my own hunch was partly disproved, and that's worth writing down.
Subscriptions — Pro, Max — are one fifth of revenue: $3.2 billion out of $16.3 billion in the first half. The other 80% is API consumption (p. 138). And the ~900% Net Revenue Retention that everybody is about to quote explicitly excludes Pro and Max subscribers, "whose fixed-price subscriptions do not reflect the expansion and contraction behavior" (p. 2–3).
The filing discloses no per-segment margin, so I can't tell you what they make on my Max plan. But there is something else in the notes that sits closer to what I'd actually noticed.
Revenue recognition for subscriptions, page F-21:
"Intelligence service provided on a subscription basis is subject to usage limits that reset over short periods, which results in revenue being recognized ratably over the subscription period."
The usage limits aren't only capacity control. They are the mechanism that decides the pattern: revenue is recognised evenly across the period rather than as it is actually consumed. "Subscription fees are generally prepaid" (p. 144). The money arrives up front and is booked in equal slices, while the limit caps the compute those slices have to cover. Even revenue, bounded cost.
Then page 29: "We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities, which could adversely affect our reputation and the perceived value of our offerings. Some of these changes have resulted, and could in the future result, in claims that our disclosures regarding model access or capabilities were inadequate or misleading."
That isn't hypothetical. The litigation note discloses that the company faces several putative class actions alleging it markets and sells Claude subscription plans "through false, misleading, and deceptive advertising practices." Anthropic disputes the allegations.
I have no view on the cases. I just note that the irritation every Max subscriber has felt in a week when they hit the limit earlier than expected is now written down as a disclosed risk factor in a securities filing.
Prices are falling, and that part isn't a hunch
I had the impression OpenAI's new model was ten times cheaper. It isn't. GPT-6.1 Terra shipped today at $1 per million input tokens and $5 per million output. Its predecessor, GPT-5.6 Terra, was $2 and $12. Call it half price — good, but not ten times.
The ten-times does exist, though. It's on my own invoice, from the other vendor:
| Same tier, one generation | Input $/M | Output $/M |
|---|---|---|
| Claude Opus 5 → Opus 5.5 | 5 → 4 | 25 → 20 |
| Claude Haiku 4.5 → Haiku 5.5 * | 1 → 0.10 | 5 → 0.50 |
| GPT-5.6 Terra → GPT-6.1 Terra | 2 → 1 | 12 → 5 |
* Haiku 5.5's $0.10/$0.50 applies to prompts under 100,000 tokens; above that it is $0.50/$2.50 — a halving rather than a tenfold cut.
One more detail pointing the same way: Sonnet 5's $2/$10 was announced as introductory pricing through 31 August 2026. The scheduled rise to $3/$15 was cancelled, and that is now the standard price (footnote on Anthropic's pricing page, checked 27 October).
Haiku dropped by a factor of ten within the same tier, in one generation. And Opus cache reads went from 50 cents to 20 per million — not only because the input price fell, but because the ratio itself changed: a cache hit costs 10% of the input price on Opus 5 and 5% on Opus 5.5.
The striking part is that Anthropic writes it into its own risk factors, on page 60:
"We may be required to reduce our prices, offer shorter contract durations or offer alternative pricing models, such as outcome or task-based pricing, any of which could adversely affect our business."
They name open source as the source of the pressure: open-weight models "may capture a portion of the market for certain use cases, which could exert pricing pressure on our offerings" (p. 41).
The same risk factor on p. 60 makes the counterpoint, and it deserves to be taken seriously — in their words, because the causal chain is theirs, not mine:
"our agentic offerings are structurally more compute-intensive than our non-agentic offerings. If we are unable to implement pricing models that adequately reflect the higher compute intensity of these agentic workloads, or if competitive dynamics constrain our ability to price these products at levels sufficient to cover their costs, our compute efficiency could deteriorate as agentic usage grows."
That is the exact direction my own usage is heading. Four parallel sessions in an assembly line is not a chat conversation.
Cost per task falls faster than cost per token
The token price is the visible half. The other half is on page 195, and it's the one that matters:
"Customers' cost per task can decrease for two reasons: newer, more capable model versions can complete tasks more efficiently, and we may reduce the prices we charge for our models."
Two effects that multiply. The filing gives both numbers for the last two model transitions — but not how they split:
| Model transition | Token price | Cost per task | Rest = efficiency (my arithmetic) |
|---|---|---|---|
| Fable 5 → Fable 5.1 | unchanged ($10/$50) | −25% typical, −45% on heavy agentic coding | all of it |
| Opus 5 → Opus 5.5 | −20% | −40% | ~25% |
That last column is mine, not theirs. The filing says explicitly that the drop is "a combination of a lower price and improved efficiency, the contributions of which depend on the task" — so the ~25% is arithmetic on their two numbers, not a figure they publish.
Look at the first row. Fable 5.1 costs exactly the same per token as Fable 5. The entire 25–45% drop is the model finishing in fewer tokens. It works like a price cut, but it was never announced as one — the token price never moved.
And the largest saving lands on agentic work — notable, because that is precisely the category the risk factors call structurally more expensive to serve.
The sentence immediately after it is the one I'd have written myself if I were selling the business:
"Because of our compute efficiency, we have been able to make these improvements while preserving our margins."
The cuts don't cost them margin. And they pay for themselves: when Opus 4.5 shipped in November 2025 with both higher capability and a lower cost per task, quarter-over-quarter revenue growth accelerated from 48% to 132% (p. 195). They credit several things at once — lower prices, higher capability unlocking new use cases, new products, more customers — so this is coincidence in time, not proof. But it points the same way: cheaper tokens didn't become less revenue. They became more tokens.
So who serves the next token?
This is what the 85% actually means. A gross margin that size on something turning into a commodity isn't a moat. It's an advertisement.
Anthropic's own lawyers wrote the argument down, better than I can. Page 40:
"competitors, customers, or other third parties have attempted in the past and will likely continue to attempt to replicate, distill, extract, or otherwise derive the functionality or capabilities of our models … These actions could … enable the development of competing products at lower cost, with reduced research and development investment."
And page 43:
"demand for frontier models may decline if customers determine that less advanced, lower-cost models are sufficient for their particular use cases … customers may conclude that incremental performance gains at the frontier do not justify the additional cost, and may migrate toward smaller, cheaper, or open-source alternatives that adequately address their needs."
I know that risk from my own work. I've run into it a couple of times this year. Some of my own development work would run fine on a model as it looked three months ago. Not the hard problems. But enough of them that if the price started to hurt, I could probably make it work without a frontier model — and with someone who hosts the weights themselves and isn't funding $9.7 billion of research per half-year.
That provider can live on a far thinner margin, because all it has to earn on is running GPUs. The 85% is what pays for the laboratory standing behind it. Two completely different businesses selling the same product.
Two caveats, so the arithmetic doesn't come out too neat. The 85% is a blended figure across prepaid subscriptions with usage caps and enterprise contracts with minimum commitments — the margin on an individual API token need not be 85%. And the commodity argument assumes three-month-old capability stays good enough. The filing bets the other way: that each frontier step opens work that wasn't possible before, and that this is what you're paying for.
Both can be true at once. The frontier model sells to whoever needs the newest, at a high margin. Everything routine drains toward whoever sells tokens cheapest. That isn't a catastrophe for Anthropic. It's just a different business from the one 85% describes today.
What I take from it
The hunch held, but not in the shape I held it. They aren't primarily making money on subscriptions — they're making it serving the people who pay — mostly API — at a blended 85% gross margin, with subscriptions as a fifth of the business whose usage limits are part of the accounting machinery.
Prices are falling, and the reason is right there in the numbers: when revenue per compute dollar climbs that fast, you can cut the price without touching the margin — they say so in those words. They've done it on Opus and by a factor of ten on Haiku, and cost per task is falling faster than cost per token, because the model also finishes sooner. OpenAI halved its prices today.
What I hadn't expected is that the same margin is the best reason to think the fall continues without them. 85% on a commodity doesn't hold once someone can deliver the same product without paying for the laboratory. I think we end up buying the hard work from frontier models and everything routine somewhere else — and I think we find out where that line sits the first time the price actually hurts.
And the 85% margin that makes it possible wasn't in the income statement. It was on page F-39, in a supplemental note.
Some days it's worth pulling the document yourself.