80 percent of Anthropic Q3 revenue came from six-month-old models

A team that pays for Claude pays mostly for models that did not exist in April. The New Stack read Anthropic's S-1. It reports that models released in the six months before September 30 drove more than 80 percent of Q3 revenue. Anthropic filed the S-1 with the SEC on Monday. It plans a Nasdaq listing under the ticker ANT.

What the filing reports about the eight models

The S-1 says Anthropic released eight widely available Claude models between April and September 2026. It also lists three restricted-access Mythos models. The filing does not break revenue down by model. Models released in late September had only days left in the quarter. The 80 percent figure covers the whole set.

Reuters reported that attendees at Anthropic's October 14 investor meeting put third-quarter revenue at $16.8 billion. The S-1 does not confirm that number. The New Stack cites both the filing and the Reuters report.

How spending ties to new releases

Anthropic says customer spending is tied to how often it releases new models. The company says each release gives existing customers more to use. New models also bring new customers in. That spending pays for the next round of research and compute. The S-1 states the pattern as the company's own account.

The filing does not say that customer demand pulled release dates forward. The two could be connected. Anthropic does not claim it in the document.

What the 80 percent does not say

The New Stack flags one gap. The 80 percent is a revenue share, not a user share. A small group of customers on the newest models could account for most of it. The S-1 also says heavy Claude Code enterprise users spend 20 times what other Claude Enterprise users spend. That gap means a revenue number cannot show how many customers switched.

The S-1 says 80 percent of Anthropic's revenue comes from consumption pricing. That number can rise or fall each month. The filing says five customers made up 25 percent of revenue in the first half. A small shift in that group changes the top line fast.

For a buyer, the point is simpler. The model running the agent today is unlikely to be the model running it next quarter. A six-month-old price list is a stale price list.

A team paying for Claude today chooses between Sonnet, Opus and a cheaper rival. The S-1 shows how fast Anthropic replaces its own catalogue. A default for a coding agent or a retrieval pipeline is often the newest, most expensive model. The S-1 shows what that default earns Anthropic back from its customers. Testing a smaller model before the next renewal is now a budget question.

Source

The New Stack on Anthropic's SEC S-1 filing.