Claude Code spikes enterprise spending: what it means for your SMB

Anthropic has filed its application to go public on the Nasdaq under the ticker ANT, and the registration document (the S-1) makes one thing clear: Claude Code is not a terminal experiment, it is the engine of its enterprise growth. For a Spanish SMB that is evaluating AI tools for its technical team, it is worth reading between the lines of what that document says.

The figure that changes everything: 20 to 1

The S-1 notes that an average Claude Enterprise user who does not use Claude Code spent less than 75 dollars a month in the third quarter. The 5% of users with the highest Claude Code consumption within Enterprise spent approximately 20 times that figure. It is not a marginal difference: it is a change in spending category.

Why? Because Claude Code does not limit itself to answering questions. It integrates into the engineer's workflow, touches the repository, proposes changes, executes tasks. And that, logically, consumes many more tokens. Anthropic describes it as a classic land-and-expand pattern: an engineer starts using it, is convinced, pushes it internally and, when the sales team arrives, the tool is already inside.

The numbers that accompany the story

The document provides several figures that should be put in context:

• More than 11,900 customers were paying 100,000 dollars or more in the last twelve months as of September 30, 2026, compared with more than 950 a year earlier.
• Revenue reached 16.3 billion dollars in the first half of 2026, compared with 1.2 billion in the same period of 2025.
• Token consumption accounted for 50% of that revenue.
• In September, Claude wrote more than 90% of the contributions to Anthropic's code, and the average engineer produced more than 45 times more code per day in the third quarter than in 2024.

Be careful with this last figure: Anthropic does not define what counts as a "contribution" or how much human review that code receives. It is an important warning for any team that wants to replicate the model without putting controls in place.

Why this matters to an SMB

If you have a small development team, the temptation is to think that all this is a matter for large corporations. Not exactly. What the S-1 describes is an adoption pattern that is also being seen in mid-sized companies: a developer discovers the tool, uses it for specific tasks and, within a few weeks, token spending skyrockets without anyone having planned it.

And there is the real risk for an SMB: the variable cost per consumption. Unlike a traditional license, here you pay per use. Without visibility or limits, a tool that starts as a trial can become an uncomfortable line on the bill at the end of the month.

There is a second, more subtle risk. Anthropic acknowledges in its risk factors that Claude Code is a critical tool in its own development, to the point that a failure of the product "could directly affect its ability to develop and deploy future models." It also warns that, as Claude takes on more engineering and research, the company "may lose visibility into certain processes in a way that is difficult for any person or group to understand and control." Translated to an SMB: if you delegate too much to the agent without traceability, you lose control of what happens in your own code.

The strategic move that almost no one looks at

The S-1 also explains why Anthropic sells the harness (the framework that runs Claude Code) to other developers. "A model alone is not an agent," says the document. Developers can run that harness on their own servers through the Agent SDK, or pay Anthropic to manage it as Claude Managed Agents. That is, Anthropic is positioning itself as infrastructure for those who build code agents on top of its models.

This connects with something we have already seen in this blog: AI agents that escape and the lessons they leave for SMBs. The difference between a useful agent and a problem is isolation, permissions and observability.

What we recommend from ForgeNEX

If you are thinking about introducing Claude Code or a similar tool into your team, these are the steps we suggest:

1. Start with a limited pilot: a couple of developers, a specific repository and a monthly spending limit defined in writing.
2. Measure consumption from day one: without token and cost metrics per project, you cannot decide whether it scales or not.
3. Define mandatory human review: the figure of 90% of contributions generated by Claude at Anthropic does not mean that no one reviews. It means that the engineer's role shifts toward direction and review.
4. Isolate the environment: minimum permissions, separate repositories for testing and traceability of every change proposed by the agent.
5. Review the data clause: confirm how the provider treats your company's code and whether it is used for training.

The S-1 also names Google and OpenAI as main competitors in frontier models, and leaves out Cursor, Replit and Cognition. It is a snapshot of the market that is worth keeping in mind: the war is not only about models, it is about who controls the layer where agents work.

For an SMB, the practical conclusion is simple: agentic AI is no longer a promise of the future, it is a line of spending and a change in the development process. Treating it as such, with budget, controls and review, is the difference between taking advantage of it and having it get out of hand. If you want to delve deeper into how Anthropic's revenue is evolving and what it implies for the market, you can read our analysis on 80% of Anthropic's revenue coming from recent models.

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