Anthropic's amended S-1: $17 billion in Q3 revenue, and what the filing tells developers building on Claude

On November 2, 2026, Anthropic filed an amended Form S-1 with the U.S. Securities and Exchange Commission, adding its third-quarter results to the IPO prospectus it first filed on October 26. The original filing only covered the first half of the year, as CNBC noted. The company plans to list its Class A shares on the Nasdaq Global Select Market under the ticker ANT; the share count and price range have not been set yet.

Most coverage today focuses on the headline numbers. For people who build products on Claude, the more useful parts are further down the document: how Anthropic makes its money, what it pays for compute, and what it says about pricing, rate limits and model turnover.

The numbers

All figures below come from the amended S-1.

Third-quarter revenue of roughly $17.0 billion is up 1,127% from the same quarter of 2025 ($1.38 billion). For the nine months ended September 30, revenue reached $33.24 billion, up 1,206% year over year.

Growth is slowing in relative terms. Revenue rose about 47% from Q2 to Q3, after 132% in Q1 and 144% in Q2, based on the quarterly table in the filing (CNBC reports the same figures). The filing itself says Anthropic expects its growth rate "to decline over time".

The net loss looks enormous, but most of it is accounting. Of the $23.29 billion Q3 loss, $21.94 billion comes from the change in fair value of convertible notes, derivative liabilities and SAFEs. The nine-month net loss is about $104 billion on the same basis. The operating line is the one to watch: Anthropic posted a small operating profit in Q2 and slipped back to a $567 million operating loss in Q3.

Most of the revenue is metered tokens

Anthropic calls its usage-based model "pay-for-intelligence": customers pay per token, priced by model. For the first nine months of 2026, consumption-based revenue was $26.3 billion (79%) and subscription revenue (Claude plans) was $6.9 billion (21%). The filing says this revenue comes primarily from customers building applications on Claude or using Anthropic's own apps, and that consumption will remain "the substantial majority" of revenue.

The customer base is also less concentrated than it was. No single customer accounted for more than 10% of nine-month revenue, though the top five together made up about 25%. In Q3, about 79% of revenue came from customers outside the top five, up from roughly 64% in Q1. Anthropic counts more than 11,900 customers paying $100,000 or more over the trailing twelve months (up from more than 950 a year earlier), more than 1,900 above $1 million and about 250 above $10 million.

Prices keep falling, and that is now visible in the revenue

For developers, the most concrete disclosure is about pricing. Between Q2 and Q3, higher usage was "partially offset by a decrease of approximately $600 million from price reductions introduced as part of new model versions and contractual discounts". The filing gives one example: Opus 5.5 cut customers' cost per task by about 40% compared with Opus 5, through a mix of lower prices and better efficiency.

Model turnover is fast. Model versions released in the six months ended September 30 accounted for more than 80% of Q3 revenue. Anthropic also warns that considerations around safety and model welfare may lead it to keep older models running longer, and that "customers who have integrated older models into their workflows may resist migration".

The practical takeaway: if your product is pinned to one Claude model, plan on migrating within months rather than years. Cost per task tends to drop when you move, but only if you actually re-test prompts and evaluations on the new version.

Compute is the constraint

Compute and infrastructure expenses rose to 58% of revenue in Q3, up from 48% in Q2. The filing explains that this includes $5.0 billion of compute bought under short-term commitments, which "currently costs significantly more" than compute procured under multi-year deals. Anthropic says it typically buys compute 12 to 18 months before it becomes available, across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.

Two risk factors are directly relevant to API users:

• Agentic workloads are expensive to serve. Multi-step reasoning, tool calls, long context and sustained sessions "consume significantly more compute per user interaction than single-turn API calls", and could grow faster than the revenue attached to them.
• Limits can change. Anthropic writes that it has "changed, and may in the future change, usage limits, rate limits, or model and product capabilities", and that some of these changes have led to claims that its disclosures were inadequate.

To secure capacity, Anthropic has signed purchase commitments and leases worth about $895.4 billion, "substantially all" related to compute and infrastructure. About $447.5 billion falls before the end of 2030, most of the rest between 2031 and 2045, and about $69.8 billion is due in the 12 months after September 30. The company had $95.6 billion in cash and $36.3 billion in short-term investments at the end of the quarter.

What builders should take from it

The filing is written for investors, but a few points carry over to engineering decisions:

1. Budget for falling unit prices, not flat ones. Anthropic is cutting cost per task with each major model and says lower prices drive more usage. Fixed-price products built on Claude can gain margin over time, provided they keep up with model upgrades.
2. Treat rate limits as a variable. Capacity is being added at high cost, and agentic usage is the heaviest load. Retries, queuing and a fallback model or provider are cheap insurance.
3. Expect seasonality. The filing notes that total customer spend fell about 20% in the two weeks ended January 2, 2026, compared with the previous two weeks, as companies froze code over the holidays.
4. The bet is on agents. Anthropic describes Claude Code as the product that "most concretely defined what autonomous AI work looks like", and says Claude Cowork's adoption outpaced Claude Code's at the same point after launch.

OpenAI, Anthropic's main rival, has not published comparable figures: according to CNBC, it confidentially filed its own prospectus in June but is signalling a 2027 listing. We first covered Anthropic's IPO plans in May, based on press reports at the time; the S-1 and its amendment now replace those reports with the company's own figures (the quarterly numbers are unaudited).