Anthropic S-1: What Claude Builders Should Change (2026)

Founden Team

1 November 2026

•

10 min read

What the IPO filing puts in writing about Claude's prices, model lifecycle and capacity, and what to change in your product

Anthropic made $16.3 billion in six months. That is its revenue for the first half of 2026, up from $1.2 billion a year earlier, according to the Form S-1 it filed with the SEC on October 26 - SEC EDGAR. It plans to list on the Nasdaq Global Select Market under the ticker ANT, with the share count and price range still blank - Anthropic. Goldman Sachs and Morgan Stanley lead the deal, and Bloomberg reported that marketing could start as soon as the week of November 9 - IPO Scoop.

Most coverage reads the S-1 as an investor story about valuation and losses. For anyone building on Claude it is something rarer: a description of your supplier written under securities law. It states how cost per task falls, how fast models turn over, how much compute Anthropic has bought, and which customer behaviours it lists as risks to its growth. This dispatch covers the numbers that matter to a product running on Claude and what to change this week.

1. The Numbers Behind the Filing

The quarterly table is the clearest picture of the business. Revenue grew from $376 million in the first quarter of 2025 to $11.56 billion in the second quarter of 2026, while compute and infrastructure costs grew from $944 million to $5.56 billion - SEC EDGAR. Compute fell from 251% of revenue to 48%, because Anthropic buys capacity 12 to 18 months ahead and revenue outran the committed cost. The first-half operating loss was $693 million; the headline $80.7 billion net loss is mostly an accounting charge on convertible notes - CNBC.

The revenue mix tells you who the customer is. Pay-for-intelligence, usage metered in tokens, brought in $13.1 billion (80%) of first-half revenue against $3.2 billion (20%) from subscriptions, and the filing says that usage comes mainly from customers building applications on Claude or using Anthropic's own apps - SEC EDGAR. In other words, the API is the business. Control does not change at the listing either: a Founder LLC of the seven co-founders holds 50.1% of the vote - CNBC. The people who set usage policy and retirement dates today keep setting them afterwards.

2. Cost per Task Falls With Each Release

This is the most useful passage for builders. With Fable 5.1 in September, Anthropic says it cut the cost of typical workloads by about 25% versus Fable 5, and by up to 45% for complex coding and agentic tasks. Opus 5.5 cut cost per task by about 40% versus Opus 5. Each cut combined lower prices with better efficiency, made "while preserving our margins" - SEC EDGAR. Opus 5.5 lists today at $4 in and $20 out per million tokens. Fable 5.1 sits at $10 and $50.

The cuts follow from how Anthropic pays for compute. Capacity bought a year ahead is a fixed cost, so the company earns most by filling it, and the way to fill it is to make each unit of work cheaper. The filing gives its own example: after Opus 4.5 launched in November 2025 at a lower price, quarterly revenue growth went from 48% to 132%, though Anthropic credits other factors too - SEC EDGAR. The skeptic's reading points the same way: Paul Kedrosky expects "the ability to charge premium prices on the latest models" to keep falling - Paul Kedrosky.

For your product, that means planning for falling costs and pricing on outcomes rather than tokens, as our guide to pricing an AI product against token costs explains. The saving only arrives when you move to the new model, and the levers you already hold still count: the S-1 puts prompt caching at up to about 97% lower cost for long prompts, which our prompt caching guide walks through.

3. Your Model Has About a Year

One figure explains the release cadence. Model versions released in the six months before September 30 brought in over 80% of revenue in the third quarter - SEC EDGAR. Kedrosky reads that as "a rapid race against churn." For a builder it means most paid Claude usage runs on models less than six months old, and a product that stays behind ends up on an older, less supported model.

Anthropic's documentation sets the floor. Current models carry a retirement date "not sooner than" about a year after launch (Opus 5.5: not before September 22, 2027), with at least 60 days' notice - Claude Docs. Sonnet 4.5 retires November 30, with Sonnet 5.5 as the named replacement, and API version 2026-10-21 removed temperature, top_p and top_k, which return a 400 error on that version. Amazon Bedrock and Google-managed endpoints set their own dates.

The S-1 adds two things the docs do not. Anthropic "may in the future change, usage limits, rate limits, or model and product capabilities," and model safety and model welfare may lead it to "limit or delay the deprecation of older models" - SEC EDGAR. Plan for a migration every six months: a fixed eval set, the model ID in one config value, and a dry run on each release. Our Opus 5.5 migration guide lists the breaking changes from the last one.

4. Capacity Is Pre-Bought, and Can Still Run Short

The S-1 lists the compute contracts behind those prices. Anthropic owes Broadcom about $161.2 billion in equipment leases, has agreed to spend up to $120.5 billion with SpaceXAI (about $3.1 billion a month once capacity arrives, expected in December 2026), and has committed at least $114.5 billion to Google, $110.0 billion to Amazon and $44.9 billion to Microsoft, those three owed regardless of usage - SEC EDGAR. With Nscale and Lambda, the agreements that state a dollar figure add up to about $637 billion, several of them maximums.

Buying ahead still does not guarantee you a response. The risk factors recall a March 2026 outage caused partly by a sudden rise in demand and compute limits, and warn that faster-than-planned demand could bring "limitations on customer and user access" - SEC EDGAR. The same Claude model IDs run on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS - Claude Docs. A second route to the same model is a failover that needs no prompt rewrite, though feature availability varies by deployment.

5. Anthropic Sells the Apps Too, and Enforces Its Terms

The filing shows Anthropic moving up the stack. Over half of its customers spending $100,000 or more a year now use both the Claude Platform and at least one Anthropic app, up from about 30% in late 2025. The top 5% of Claude Enterprise users of Claude Code spend about 20 times the under-$75 monthly average of non-users, and Claude Managed Agents runs the whole agent harness on Anthropic's compute - SEC EDGAR. The commercial terms let you power products for your own customers, but not "build a competing product or service" or resell without approval - Anthropic.

Enforcement is now material. The S-1 estimates that 5% to 7% of revenue in the nine months to September came from customers later found violating the usage policy or commercial terms, and says enforcement tightened in August and September 2026 - SEC EDGAR. The most telling line sits in the risk factors: customers may hold back because of a "preference for multi-vendor AI strategies" to limit dependence on one provider. Anthropic names that hedge as a risk to its own revenue.

The portable layer is getting easier to build on. On October 22, Anthropic donated Agent Skills (about 1 million published skills, nearly 280 million installs) to the Linux Foundation's Agentic AI Foundation, which already hosts MCP - Anthropic. Skills, MCP tools and one shared AGENTS.md now carry over between AI coders. Founden made a similar move on October 1, when its desktop app added Google's Antigravity as a third AI coder alongside Claude Code and Codex.

6. What to Change This Week

The S-1 does not change how Claude works today. It changes what you can count on: costs fall per release, models turn over yearly, capacity is finite, and the supplier competes in the app layer while enforcing its terms more strictly. Each is something to plan around, and each plan costs far less than a forced migration or an outage.

Start with the deadline. Move off Sonnet 4.5 before November 30, and check that no request still sends temperature, top_p or top_k. Then re-run your evals on Opus 5.5 and on Sonnet 5.5 at $2 per million input tokens, and record cost per task rather than price per token, because that is the number the S-1 says keeps falling.

Then add resilience. Set up a second route to the same model through a cloud provider and test the failover. Check your use against the commercial terms, especially if you resell access. And keep a second lab tested: OpenAI's GPT-6.1 Terra lists at $1 per million input tokens, and model routing can split work between labs.

The pattern behind these steps is that the model is a supplier, not a foundation. The S-1 confirms that Claude will keep getting cheaper and better, and that it will keep changing on Anthropic's schedule. Products that keep the model ID in one setting, an eval set ready, skills and tools in open formats and a second route warm get every price cut without the scramble. That is control over the model, not immunity from the lab's decisions, and it is cheap to put in place before the next release.

This dispatch reflects Anthropic's Form S-1 as filed on October 26, 2026. The registration statement is not yet effective, and figures can change in amendments before the listing. Model prices and retirement dates change often, so check the current documentation before you act.

What this article rests on

Every model figure linked above is read from its record, re-read from OpenRouter's model feed every 6 hours, so the figures on this page are current even where the prose around them is older. Prices are OpenRouter's listed price per million tokens: for a lab's closed models that is the lab's own list price, and for an open-weight model it is the price of the host OpenRouter routes to by default, which can differ from the lab's own API and from other hosts.

Claude Opus 5.5 (Anthropic): $4 in and $20 out per million tokens, cached input $0.20, a 1M token window, listed 22 September 2026.
Claude Fable 5.1 (Anthropic): $10 in and $50 out per million tokens, cached input $0.25, a 1M token window, listed 1 September 2026.
Claude Sonnet 5.5 (Anthropic): $2 in and $10 out per million tokens, cached input $0.10, a 1M token window, listed 28 September 2026.
GPT-6.1 Terra (OpenAI): $1 in and $5 out per million tokens, cached input $0.05, a 1.05M token window, listed 26 October 2026.

In this topic: AI models

Which AI model should you build on? Every guide on it

Data behind this topic: AI model prices, context windows and release dates.