Anthropic's Amended S-1: What API Customers Should Know

On November 2, 2026, Anthropic filed Amendment No. 1 to its Form S-1 registration statement with the US Securities and Exchange Commission, one week after the original filing on October 26. A registration statement is a company's own description of its business, written for prospective investors and reviewed by a regulator. It is not marketing, and it is more candid than any launch post or sales deck about how it sells, what it has promised to spend, and what could go wrong.

We read both documents on EDGAR and compared them. This post is written for the people who buy Claude rather than the people who might buy shares: engineering leads, procurement teams and operators whose products depend on the API or on Claude through Amazon Web Services, Google Cloud or Microsoft Azure. Every number is company-reported, every section is named, and the filing is not yet effective. Where the document is blank, we say so rather than fill the blank from press coverage.

Three things are deliberately absent. The filing states no valuation, no share price range and no offering date, so none appear here. We offer no view on the shares. And we do not link to our own June commentary on reports of a confidential draft; the public filings supersede anything written before October 26.

Two documents sit on EDGAR under Anthropic's Central Index Key 2130357. The original Form S-1 was filed on October 26, 2026. The S-1/A, Amendment No. 1, followed on November 2, 2026 under registration number 333-299633. Both state that Anthropic has applied to list Class A common stock on the Nasdaq Global Select Market. Both leave the price range as blanks on the cover.

Why a buyer should care about the form at all: a registration statement carries liability for material misstatement, so its risk factors are written to be complete rather than reassuring. When a vendor's own lawyers list the ways its service can fall short, that list is a better vendor-risk input than any sales deck. The sections quoted below are Prospectus Summary, Risk Factors, Management's Discussion and Analysis (MD&A), Business, Use of Proceeds and the notes to the financial statements.

One caution on reading it. The filing defines a metric called Run-Rate Revenue and refers to it repeatedly, but leaves its value blank as of October 31, 2026 in both documents. Any article quoting an Anthropic run-rate figure as of that date is not quoting the filing.

We compared the two documents word by word. Most of the text is identical. The substantive changes fall into five groups, and the table lists each with the before and after as the filings state them.

The example that shows why the diff matters: a buyer reading only the October document would have known that Anthropic had large compute commitments but not their size. The amendment puts a number on them, and the number reframes every capacity sentence in the risk factors. The failure case is the opposite reading, where someone treats the net loss as an operating result. The filing is explicit that the $104.0 billion is primarily a fair-value adjustment on convertible instruments, while the operating loss for the nine months was $1.3 billion, down from $5.2 billion a year earlier, and the company's own Adjusted Operating Income measure was $4.1 billion.

The Risk Factors section states the annual sequence plainly: revenue of $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025, which the company describes as growth of 841% and 1,088% in the two later years. For the nine months to September 30, 2026, revenue was $33.2 billion against $2.5 billion in the same period of 2025. The Prospectus Summary adds that quarterly revenue went from zero at commercial launch in the first quarter of 2023 to $17.0 billion in the third quarter of 2026.

The mix is the part a buyer should read twice. The MD&A states that consumption-based revenue, the usage metered in tokens and priced by model, was $26.3 billion or 79% of the nine-month total, with subscriptions at 21%. The filing also notes that consumption revenue was 75% of the total in 2024 and 83% in 2025. Anthropic is, by its own accounting, overwhelmingly an API and usage business, and the risk factors that follow are written with that mix in mind.

The company also warns, in the same section, that it expects its Run-Rate Revenue growth rate to decline over time and that the decline may be significant and faster than it anticipates. That is standard language in a growth-stage filing, but it is worth noticing that it is attached to the one metric left blank.

For a buyer, concentration cuts two ways. A vendor dependent on a handful of customers may bend its roadmap toward them; a vendor with a broad base is steadier but has less reason to accommodate any one account. The filing gives enough detail to see which way Anthropic has moved.

The Business section states that customers generating $100,000 or more of trailing-twelve-month revenue grew from more than 950 at September 30, 2025 to more than 11,900 a year later. The MD&A adds that revenue from those customers was $24.5 billion in the nine months to September 30, 2026, about 74% of the total. The company reports Net Revenue Retention of approximately 900% as of the same date. The filing computes it as qualifying customers' revenue in the three months to the measurement date divided by their revenue in the same three months a year earlier, excluding Pro, Max and Ultra subscribers.

At the top of the distribution, the notes to the financial statements list customers at 10% or more of revenue. In 2023 one customer was 43% and another 11%. In 2025 two customers were 12% each. For the nine months to September 30, 2026, none reached 10%. The Business section puts the same point differently: about 79% of third-quarter 2026 revenue came from outside the top five customers, up from about 64% in the first quarter. The Risk Factors add that the top five customers together still represented approximately 25% of revenue for the nine months to September 30, 2026.

One sentence in the Business section is aimed squarely at enterprise buyers deciding how to deploy. In the three months to September 30, 2026, Claude Enterprise users who did not use Claude Code spent on average less than $75 a month, while the top 5% of Enterprise users who did use Claude Code spent approximately twenty times that. It is a sales argument, but it is also a budgeting warning: agentic coding usage is where the spend concentrates.

The channel concentration is the number most buyers will not have seen. The Risk Factors state that revenue through cloud service provider marketplaces was about 11% of total in 2023, 32% in 2024, 47% in 2025 and 44% in the nine months of 2026, with net accounts receivable through those marketplaces at 52% on September 30, 2026. Nearly half of what Anthropic sells is invoiced through Amazon, Google or Microsoft.

The sentence the amendment adds is long and worth paraphrasing carefully. Anthropic states that it has entered into purchase commitments and lease agreements, including short-term commitments and leases not yet commenced, with aggregate committed spend of approximately $895.4 billion, substantially all of it related to compute and infrastructure. Of that, approximately $447.5 billion is expected to be incurred before December 31, 2030, and approximately $69.8 billion falls due in the twelve months after September 30, 2026. The figure appears in the Prospectus Summary, the Risk Factors and the MD&A's contractual obligations discussion.

The Risk Factors describe the terms: minimum purchase commitments and take-or-pay provisions that obligate the company to pay for resources regardless of actual utilisation. The architecture behind the spend is named as a multi-chip approach across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. The suppliers are named too: a limited number of cloud providers, listed as Amazon, Google and Microsoft, plus suppliers such as Broadcom. Amazon and Google also hold convertible instruments in the company, with Amazon's conversion capped at 12% of capitalisation.

Why this matters to an API customer: those commitments are the capacity your requests run on. A take-or-pay contract gives Anthropic a strong reason to keep utilisation high, which is good for availability in ordinary times and a reason to expect tight capacity management when demand spikes. The example the filing itself gives is March 2026, covered in the next section. The failure case for a buyer is assuming that a vendor with $895 billion of committed compute cannot run short; the filing says the opposite can happen when demand accelerates faster than compute that was bought in advance.

For scale, the filing states that Anthropic has raised an aggregate of $141.1 billion of capital and held $95.6 billion in cash and $36.3 billion in short-term investments, with total indebtedness of $6.2 billion. The Use of Proceeds section says offering proceeds will go to general corporate purposes including compute and infrastructure, with the amounts blank.

This is the section an API buyer should copy into their vendor-risk file. The Risk Factors state that in March 2026 Anthropic experienced a service outage and periods of degraded performance, attributed to a combination of factors including a significant and rapid increase in usage demand and compute constraints in certain components of its infrastructure. We wrote about that incident at the time in our March 2026 outage resilience playbook; the filing now confirms the cause in the company's own words.

The structural statement is more useful than the incident. Because compute is procured well in advance based on projected demand across all workloads, the filing says, a significant acceleration in customer demand could cause aggregate capacity to fall short of what is needed to support all workloads concurrently. Read with the compute section above, this describes a supply curve that steps up on contract dates rather than one that flexes with your traffic.

Then the sentence that bears directly on integration design. Anthropic states that it has changed, and may in the future change, usage limits, rate limits, or model and product capabilities, and that doing so could harm its reputation and the perceived value of its offerings. Elsewhere the filing notes that subscriptions are subject to usage limits that reset over short periods. None of this is new to anyone who has run production traffic on the API, but it is now a disclosed risk rather than an anecdote.

Two further operational risks are named. Utility outages, shortages, price increases or required curtailment at cloud data centres could affect service, and the company notes it has publicly committed to covering the cost of electricity price increases to consumers caused by its data centres. And if any of the three cloud providers raised prices, terminated the relationship or reinterpreted its terms of service, the filing says that could degrade the availability and performance of Anthropic's models. The practical response is the one we set out in our provider failover and safe-retry guide: a second route to the same model family, and idempotent retries that do not amplify an incident.

Buyers tend to ask two forward-looking questions: will the price fall, and will my model be retired. The filing answers neither directly, but it frames both.

On price, the Risk Factors treat open-weight models as a competitive pressure that could push pricing down or move some customers to alternatives. The filing does not say Anthropic intends to cut prices, and it does not say it intends to hold them. What it does say, in the MD&A, is that customers pay for usage metered in tokens and priced according to the model selected. That is the pricing model our Fable 5.1 cost guide works through line by line, and nothing in the filing changes it.

On deprecation, two sentences stand out. First, the company writes that considerations relating to model safety, interpretability, alignment and welfare may lead it to limit or delay the deprecation of older models, or to continue supporting legacy versions alongside newer releases, creating trade-offs in product development and compute distribution. Second, it acknowledges that customers who have integrated older models may resist migration, and that accelerated or forced deprecation could result in attrition, contractual disputes or reputational harm.

The example of how to use this: if you are negotiating an enterprise agreement, the filing hands you the vendor's own statement that forced deprecation is a recognised harm. Ask for notice periods in writing. The failure case is reading the welfare sentence as a promise that old models stay available; it is a statement of tension, and the same paragraph says the trade-off lands on compute distribution. Our model alias and retirement ledger tracks the actual dates as vendors publish them.

A filing is evidence, not a forecast. The useful move is to update specific assumptions in a vendor-risk register rather than to form a view on the company. Four situations cover most of the teams our AI transformation practice advises.