Notes on coding agents and the pages they leave behind, on moving between notes apps, and on the plain mechanics of keeping what you think. Written and published from a Sumibako vault.
Anthropic's IPO filing, read by a Claude user: what the S-1 says about limits, outages and Claude Code
Anthropic filed its S-1 on 26 October. What it tells people who pay for Claude: usage limits as a risk, the March outage, Claude Code, and the numbers.
Prashant Shaw
Founder, Sumibako
8 min read
On 26 October Anthropic filed its S-1 with the SEC, the document a company publishes before it sells shares to the public. It has applied to list on the Nasdaq Global Select Market under the ticker ANT. The price range is still a blank line, as is the number of shares, so nobody can tell you yet what a share will cost.
I did not read it as an investor. I read it as somebody who pays for Claude, builds on Claude Code, and has hit the usage limit at the worst possible moment more than once. A prospectus is the one document a company has to write carefully, because it can be sued over every sentence, and this one says a surprising amount about the product you use every day. Everything below is quoted from the filing, checked on 28 October 2026.
❗IMPORTANT
This is not investment advice. I run a small notes app and can barely be trusted with a usage meter.
📄 The numbers, before the reading
Most of the coverage this week has been about these, so here they are in one place, straight from the summary and quarterly tables.
The line that looks like a typo is the net loss for the first half of 2026: $80.7 billion. It is not a typo. Nine tenths of it, $72.4 billion, is a change in the fair value of convertible notes, which the filing explains were "primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation". In plain words, the more the company is worth, the more it loses on paper. I would like that arrangement for my own accounts, but only in that direction.
The filing also says it had raised "an aggregate of $141.1 billion of capital" by 30 June, and that revenue "grew 1,368% year over year for the three months ended June 30, 2026 compared to compute and infrastructure expenses that grew 300% during the same period." That second sentence is the one to remember when you read the next section.
🧮 Where a $20 subscription fits
If you pay for Pro or Max, you are not the main event. "Pay-for-intelligence, our consumption-based revenue model, is our primary revenue model and currently represents the substantial majority of our total revenue." That is the API and usage-based enterprise contracts. Subscriptions are described separately, and their defining feature is the one you already know: they "are subject to usage limits that reset over short periods".
The filing even explains what the limit is for, from the company's side of the counter: "when subscribers approach or exhaust the usage capacity included in their tier, they may utilize a higher subscription tier or purchase more intelligence directly, creating a seamless bridge from subscription to incremental pay-for-intelligence revenue." I have met that seamless bridge. It said I could continue at 2pm.
One more detail for subscribers. The headline growth metric, a Net Revenue Retention of "approximately 900%", leaves you out on purpose, "excluding subscribers to our Claude Pro and Max plans", because a fixed price cannot expand. That is fair, since a fixed price is the point of a subscription. It also tells you which customers the growth story is about.
⏳ Usage limits are now a risk factor
This is the paragraph I would put in front of every Claude user. Among the risks it has to disclose, Anthropic wrote:
"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 is a company telling its future shareholders, under legal obligation, that the limits you work within are a lever it has pulled before and may pull again. It is not a scandal. Every subscription with a fair-use clause works this way, and I have written about how Claude's limits behave today. But it is unusually plain, and it sits next to another one about capacity: if demand outruns the compute bought for it, "we may experience service disruptions, degraded response times, or limitations on customer and user access to our models, platform, and applications".
It has happened. "In March 2026, we experienced a service outage and periods of degraded performance resulting from a combination of factors, including a significant and rapid increase in usage demand, compute constraints in certain components of our infrastructure, and other factors." The fix it describes is reasonable, more scalable infrastructure, and so is the honesty of the next line: "similar incidents may occur in the future."
🧑💻 What it says about Claude Code
Claude Code gets the most enthusiastic prose in the document. It is "the clearest manifestation of Virtual Collaborators today", a phrase the filing uses for agents that do real work, and the adoption story is the one most of us lived: it "began with individual software developers who noticed the potential of autonomous coding agents before enterprise procurement teams had considered the category".
Three numbers stood out.
1️⃣ Inside Anthropic, Claude writes the code. "Over 90% of the contributions to our codebase in September 2026 were written by Claude." The average daily volume of code per engineer in the third quarter "was over 45 times the average daily volume in the year ended December 31, 2024."
2️⃣ Claude Code users spend more, by a lot. "Claude Enterprise users who did not use Claude Code spent less than $75 per month on average, while the top 5% of Claude Enterprise users who used Claude Code spent approximately 20 times that amount." If you have watched an agent read your whole repository twice before lunch, this will not surprise you.
3️⃣ Newer models got cheaper per task. With Opus 5.5 "we reduced customers' cost per task by approximately 40% compared to Opus 5", through a mix of lower prices and fewer tokens per job. Fable 5.1 cut "the cost of typical workloads by approximately 25%".
The filing is also frank about misuse. In September 2025, it says, "a Chinese state-sponsored hacking group manipulated our Claude Code tool into attempting infiltration into roughly 30 global targets", which it calls "the first documented case of a large-scale cyberattack executed without substantial human intervention."
🚫 The weeks a model was switched off
Two government actions in the risk factors touch users directly.
The first is the one you may have noticed in June. "The U.S. Department of Commerce imposed worldwide export restrictions on our Fable 5 and Mythos 5 models. To ensure compliance, we disabled Fable 5 and Mythos 5 for all of our customers; access to our other models was not affected." The models came back, in the filing's words, "after we updated our safeguards in coordination with the U.S. government and our industry partners and once the U.S. Department of Commerce lifted its restrictions." The filing warns that "Similar export controls could recur, could remain in place longer, or could extend to additional models or capabilities."
The second is the dispute with the Department of War, which in February moved to designate Anthropic a "supply-chain risk to national security" after the company said publicly that its models should not be used for mass domestic surveillance or fully autonomous weapons. Anthropic largely won in the district court on 27 August. On 25 September the D.C. Circuit, "in a 2-1 split decision", upheld one of the designations, and Anthropic has asked the full court to rehear it. The filing says it "cannot predict the ultimate outcome".
For an ordinary user, the lesson of both is narrower than the politics: the model you rely on can become unavailable for reasons that have nothing to do with you or with Anthropic's engineering. Keep a second model in mind for the work that cannot wait.
⚖️ Lawsuits a Claude user might recognise
The authors' settlement. The copyright class action known as Bartz settled "for $1.5 billion, which was approved by the district court in July 2026." In 2025 it has a line of its own in the income statement, bigger than everything spent on sales and marketing that year.
Music publishers and Reddit have pending suits over lyrics and scraped posts.
A new one, for subscribers. In September a putative class action accused Anthropic and other frontier labs of agreeing "to restrain competition by slowing the pace at which we develop and improve our respective AI models", on behalf of "a purported class of purchasers of the defendants' paid consumer subscriptions." If you pay for Pro, you are in the class it describes. The claim, in short, is that the models are improving too slowly. Anthropic says it disagrees and will defend it. Whatever happens, it is the most 2026 lawsuit I have read.
🏛️ Who will be in charge afterwards
Shareholders will not run it in the usual way. Anthropic is a public benefit corporation, and a body called the Long-Term Benefit Trust, through a special class of stock, "generally has the authority to elect and remove the Class T directors, who together will constitute a majority of the authorized members of our board of directors." The founders get a single Class F share through a company they control, and Anthropic will be a "controlled company" under Nasdaq's rules. Whether that is reassuring depends on whether you trust the trustees more than the market. The filing at least spells it out.
🗂️ What I take from it as a user
Nothing in the S-1 should make anyone stop using Claude. It is the most capable tool I use, and the filing describes a company that is, for the first time, making more from it than it spends on running it. What the document does is put in writing three things heavy users already suspected: limits are adjustable and have been adjusted, capacity can run short, and a model can be switched off by a government.
So the habit I would keep is the boring one. Keep what matters outside the conversation: the plan, the decisions, the summary of a long session before it grows too long. I make Sumibako for exactly that: Claude can write a page into it through a connector on every plan, Claude Code can file one with a single command, and the page stays yours if the session, the limit or the model goes away. Sumibako is not going public. Its share price is a blank line too, for more permanent reasons.
📚 Sources
Anthropic, PBC, Form S-1, filed with the SEC on 26 October 2026: the prospectus summary, risk factors, management's discussion and analysis, and the quarterly results table, from which the figures and the chart above are taken
PHP2Shell, Railway, Next.js, LiteLLM, eMudhra, Plex, GitLab, Cairn: 13 to 24 October 2026 in order, what the numbers mean, and what the week had in common.
