80% of Anthropic's revenue comes from recent models: what it means

Anthropic has filed its long-awaited S-1 form with the SEC, the step prior to its IPO on the Nasdaq under the ticker ANT. Beyond the financial headline, there is a data point that should interest any IT team that is incorporating generative AI into its processes: more than 80% of third-quarter revenue came from models launched in the previous six months, that is, from products that practically did not exist at the beginning of the year.

Between April and September 2026, the company released eight Claude models for general access, in addition to three Mythos models with restricted access. The document does not break down how much each one billed, and it is worth remembering that those launched at the end of September had only a few days to contribute to the quarter. Even so, the conclusion is clear: Anthropic's business rests on constant catalog renewal.

A self-reinforcing cycle

The S-1 itself describes the mechanism: each model improvement gives customers more reasons to use Claude, attracts new users, and encourages existing ones to spend more. That spending funds the next round of research, computing power, and talent, which in turn produces more capable models. It is a circle in which each generation pays for the next.

What the document does not clarify is whether customer demand has pushed Anthropic to accelerate its release schedule in 2026, nor how many customers have actually migrated to the new models. A small group of accounts with high spending could explain much of that 80% figure. As a reference, the S-1 itself notes that the most intensive enterprise users of Claude Code spend approximately 20 times more than the rest of corporate users.

Reuters, citing attendees at the October 14 investor meeting, put third-quarter revenue at 16.8 billion dollars. It is a figure that does not appear in the S-1 and should be taken with the usual caution of leaks.

What it means for a small business

The quick reading would be to think that you always have to hire the most powerful and most recent model. The useful reading is different: the pace of renewal in the AI catalog is so high that the decision of which model to use must be reviewed periodically, not made once and forgotten. If your team built an assistant or an automation flow a year ago with a specific model, it is likely that today there is a cheaper alternative that does the same thing, or a more capable one that solves what it previously could not.

At ForgeNEX we see this pattern daily in small businesses that are starting to automate with AI. The temptation to go directly to the top-of-the-range model without measuring is frequent, and spending accumulates without anyone reviewing it. As we discussed in our analysis of Claude, GPT, and Gemini, the choice depends much more on the specific task than on the position in the ranking.

Practical recommendations

• Measure before scaling. Before upgrading to a higher model, check with real data from your business whether the improvement in results justifies the cost jump.
• Try the small model first. For email classification, invoice data extraction, or first-level responses, a lightweight model is usually enough.
• Review the catalog every quarter. Prices and capabilities change quickly; a periodic review prevents paying too much out of inertia.
• Set an AI consumption budget. With agents and automations, variable spending can skyrocket if there are no limits or alerts.
• Document why you chose each model. When the next launch comes, you will have criteria to decide whether to migrate or not.

This approach fits with what we already discussed about process automation with n8n and AI: the tool is important, but what makes the difference is designing the flow with judgment and reviewing it with data.

The provider's incentive

There is a nuance worth keeping in mind when reading these figures: Anthropic has a clear financial incentive to keep improving and launching models, because each new release drives customer spending. That does not invalidate the quality of the product, but it does invite your team to make the adoption decision with its own criteria and not only because of the manufacturer's push.

The company has also publicly defended the idea of moderating the pace of frontier AI development so that safety advances at the same pace as capabilities. The S-1 does not clarify whether commercial demand has conditioned that approach. It is a debate we will continue to follow closely, just as we follow the lessons on AI agents that escape.

In summary: the news confirms that the AI market is moving at a brutal pace of renewal. For a small business, the sensible response is not to chase every launch, but to set up a lightweight evaluation process that allows you to take advantage of improvements without skyrocketing spending or complexity.

Source: The New Stack. Analysis and adaptation: ForgeNEX.