What happened
On 2026-10-26 Anthropic, PBC filed a registration statement on Form S-1 with the SEC. It plans to list Class A common stock on the Nasdaq Global Select Market under the symbol "ANT". The price range has "not yet been determined".
The filing itself is long, and most of the posts about it pick one line from the financial tables and stop there. This post goes the other way: it reads the filing for the parts a developer who builds with Claude every day would want to know, and keeps the money to what the table actually says.
One thing first. The filing is not finished. Its Run-Rate Revenue sentence still reads "our Run-Rate Revenue was $ as of October 31, 2026", with the number left blank. When I checked EDGAR on 2026-10-28 there was no amended filing yet.
The money, read from the table
The headline people are sharing is the net loss: $80.7 billion for the six months ended June 30, 2026. That number is in the filing. So is the line that explains most of it: $(72,388) million of it is "change in fair value of convertible notes, derivative liabilities, and SAFEs". That is an accounting change in what earlier investors' paper is worth, not money spent running Claude.
The operating side looks very different. For the same six months:
- Revenue was $16,286 million, against $1,163 million a year earlier.
- Total operating expenses were $16,979 million, so the operating loss was $(693) million.
- Compute and infrastructure, the biggest cost, was $8,945 million.
- 80% of revenue ($13.1 billion) was consumption-based: paid per use rather than through subscriptions.
For the whole of 2025, revenue was $4,585 million, up from $386 million in 2024.
So the honest one-line summary is: revenue is growing very fast, the business is close to covering its operating costs, and the giant loss is mostly about how convertible notes are valued. Whether that holds is a question for the people pricing the shares. For a developer, the more useful parts are elsewhere in the document.
What it says about Claude Code
The filing makes claims about how Anthropic builds software, and because it is a legal document they come with risk factors attached.
The claim: "over 90% of the contributions to our codebase in September 2026 were written by Claude." It also says the average daily volume of code per engineer, in the three months to September 30, 2026, was "over 45 times" the average for the year 2024.
It describes Claude Code as a coding agent that works "on end-to-end software development: writing, planning, editing, testing, debugging, and committing code". And it gives an internal benchmark: Claude's success rate on open-ended engineering problems "reached approximately 90% in September 2026, up approximately 50 percentage points in nine months".
Then the risk factors say what that dependence costs. Claude Code is a "critical tool" in Anthropic's own development, so "any errors, vulnerabilities, or degradation in these products could directly impair our ability to develop and deploy future models and products." And this one is worth reading twice:
AI-generated code could incorporate or reproduce open-source software such that the associated license obligations, including copyleft requirements to disclose or license proprietary source code, apply to us without our knowledge.
The company that makes the coding agent is telling its future shareholders that code written by the agent might carry licence duties nobody noticed. The filing adds that similar concerns "could also lead enterprise customers to limit their use of our products."
There is also the misuse case. The filing says that in September 2025 a Chinese state-sponsored group manipulated Claude Code "into attempting infiltration into roughly 30 global targets", with "a limited number of successful intrusions".
What it says about MCP
If you build MCP servers, one paragraph matters. The filing describes MCP as "an open standard we introduced in November 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025." It lists MCP next to the Agent SDK, Claude Managed Agents and agent skills as "composition primitives" that combine into production workflows.
In an IPO document, a company lists what it owns and what could hurt it. MCP sits in the product description as something Anthropic gave away. For anyone wiring a web app to tools through MCP, that is the reassuring reading: the protocol is not a single company's asset in the filing's own words.
What it says about the cost of a task
The business model section has one sentence a web developer can use when choosing a model: "Claude Opus 5.5 reduced customer costs per task by approximately 40% compared to Claude Opus 5." The argument that follows is that cheaper tasks bring more usage.
Note the unit. It is cost per task, not price per token. A model that costs more per token but finishes in fewer steps can be the cheaper one. If your app runs an agent loop, the number to measure is what one finished task costs you, not the price on the pricing page.
The filing also says customers paying $100,000 or more a year grew from more than 950 to more than 11,900 between September 30, 2025 and September 30, 2026, and that Net Revenue Retention was approximately 900%. Those are big numbers, and the next section is about how far to trust them.
What is not known yet
- Run-Rate Revenue, the price range and the offering size. Blank in the filing as filed.
- How "contributions" and "success rate" are measured. The filing says these are internal. Its own risk factor says Run-Rate Revenue, Net Revenue Retention, customer counts and "Claude's success rate" "have not been independently verified by third parties".
- What changes before pricing. An amended S-1 can add numbers or change wording; none existed when I checked.
My take
This is my view, not the filing's.
The line I keep coming back to is not the loss. It is the copyleft risk factor. Most of the commits in the repo that runs my own posting automation carry a Claude co-author line, so this is my risk too, and the risk of anyone shipping code an agent wrote. Anthropic is saying, in the one document where understating a risk has legal teeth, that nobody fully knows what an agent copied. I read that as an instruction: licence scanning belongs in the pipeline next to the type check, not in a someday list.
The second thing is the cost-per-task framing. Price per token is what the pricing page shows, so it is what most of us compare. The filing measures something else, and for agent work I think it is the right thing to measure. And I note, without reading too much into it, that the most impressive engineering numbers in the filing are the ones the filing itself says nobody outside the company has checked.
