Anthropic’s S-1, filed with the US Securities and Exchange Commission on October 26, 2026, tells investors that safety judgments can delay new Claude models, restrict who gets them, and slow the retirement of old ones. The same risk factors say the company has already changed usage limits, rate limits and model capabilities, and may do so again.
The Anthropic S-1 Says Safety Measures Have Already Delayed Model Releases
Anthropic’s S-1 states that its precautionary safety, alignment and security measures have in the past included delaying model releases and limiting their initial availability. The example it gives is Mythos, which Anthropic released in April 2026 as a gated research preview through Project Glasswing, open to selected organizations for cybersecurity work, because of the model’s ability to find software vulnerabilities. The filing says Anthropic expects future models with similar capabilities may face access constraints too. It also says it cannot predict when such measures will be needed, which ones, or how long they will last, and that they may become “more frequent and more costly” as capabilities advance. Elsewhere the S-1 acknowledges that organizations left out of gated programs may see the access decisions as arbitrary or anticompetitive.
Model Welfare Can Slow Deprecation of Older Claude Models
The S-1 says considerations of model safety, interpretability, alignment and welfare may lead Anthropic to limit or delay the deprecation of older models, or to keep legacy versions running alongside new releases. The filing presents this as a cost to Anthropic, because legacy models consume compute and slow product work. It also describes the opposite pressure. Customers who built older models into their workflows may resist migrating, and Anthropic says a forced or accelerated deprecation could lead to customer attrition, contractual disputes or reputational harm. In the same passage, the company says it has changed usage limits, rate limits, and model and product capabilities, and that some of those changes produced claims that its disclosures about access or capabilities were inadequate or misleading.
Claude Revenue Is Mostly Consumption-Based While Compute Commitments Run for Years
Anthropic’s revenue is primarily consumption-based and much of it is not under long-term contract, the S-1 says, while its compute and infrastructure commitments generally run for multiple years. Revenue for the six months to June 30, 2026 was $16.3 billion, up from $1.2 billion a year earlier, against a net loss of $80.7 billion for the same six months. The top five customers accounted for about 25% of revenue in the nine months to September 30, and no single customer exceeded 10%. The filing says many top customers have no long-term contract or have contracts expiring soon, and that some large customers negotiated flexibility on financial commitments, termination rights, indemnification and security obligations. As of September 30, Anthropic counted more than 11,900 customers paying $100,000 or more a year, up from more than 950 a year earlier. In the third quarter, over half of them used both Claude Platform and at least one first-party application such as Claude Code, up from about 30% at the end of 2025.
The Filing Names Amazon, Google and Microsoft as Partners and Competitors
The S-1 says the three cloud providers that sell Claude are also Anthropic’s investors, customers and competitors. It warns that they may put their own compute needs, or those of rivals they have invested in, ahead of Anthropic’s, and may make platform decisions that disadvantage Claude against their own offerings. For a customer buying Claude through AWS, Google Cloud or Azure, that means the capacity behind its workloads depends on a supplier relationship that Anthropic itself lists as a risk.
Why This Matters for Engineering Procurement
The S-1 puts in a securities filing how Anthropic may change the models its customers depend on, and a buyer should let it shape the contract and the architecture together. On the contract side, the filing confirms that large buyers have negotiated terms on termination, indemnification and security. Ask for the same attention on model lifecycle: a written notice period before a model version is deprecated, notice before usage or rate limits change, and a commitment on how long a pinned version stays available. On the architecture side, assume release dates can slip and access can be gated. Do not schedule a product launch on a model that is not yet generally available, and keep an evaluation suite that can qualify a replacement model within days. The rising share of customers on both the platform and first-party apps is worth tracking inside your own company too, because every extra Claude product a team adopts raises the cost of moving later. We covered one company’s public account of switching model suppliers over cost and performance in Replit’s stated move from Anthropic to OpenAI.
FAQ
When did Anthropic file its S-1?
Anthropic filed its S-1 registration statement with the US Securities and Exchange Commission on October 26, 2026. The BBC reports it plans to list on the Nasdaq under the symbol ANT.
Can safety reviews delay new Claude models?
Yes, according to the S-1. Anthropic says precautionary measures have already included delaying releases and limiting initial availability, as with Mythos through Project Glasswing in April 2026.
Will Anthropic keep older Claude models available?
The S-1 says safety, interpretability, alignment and model welfare considerations may lead it to delay deprecating older models or keep legacy versions running alongside new ones. It gives no fixed schedule.
Has Anthropic changed Claude usage or rate limits before?
Yes. The S-1 says Anthropic has changed usage limits, rate limits and model capabilities, and that some changes drew claims its disclosures were inadequate or misleading.
How concentrated is Anthropic’s revenue?
Its top five customers made up about 25% of revenue in the nine months to September 30, 2026, and no single customer exceeded 10%. Many top customers have no long-term contract.