Key Points

• Price cuts and customer discounts resulted in a reduction of approximately $600 million in Anthropic's revenue growth in the third quarter.
• Computing and infrastructure expenditures increased by approximately 77%, which is higher than the 47% revenue growth rate.
• The company has turned to operating losses, but short-term computing power costs and a non-cash expense have made the situation even more complicated.

Anthropic is making the usage cost of Claude lower. However, its calculation bills are changing in the opposite direction.

According to the SEC document revised from Anthropic, the company stated that the revenue growth due to increased usage was "partially offset by price reductions after the launch of the new model version and contract discounts, with the offset amount being approximately $600 million."

At the same time, spending on computing and infrastructure soared from $5.56 billion to $9.86 billion, an increase of about 77%.

This means that for every $1 in revenue generated by Anthropic, approximately 58 cents are spent on computing and infrastructure, compared to 48 cents in the previous quarter. This metric includes the computing power used for research, training, and internal operations, not just the computing power used to provide services to paying Claude customers. Therefore, it is not a profit margin calculated per individual prompt.

These $600 million did not “disappear” from the accounts of Anthropic. What the company stated is that price reductions and discounts offset the revenue growth driven by increased usage.

The cost of being cheaper AI

These price reductions are not assumptions. Claude Opus 5.5 was recently released, with its price lowered by token compared to Opus 5. However, its capabilities are superior to any model before it, and sometimes it can even be comparable to Fable. In October, Anthropic also launched Haiku 5.5, positioning it as a cheaper model suitable for high-throughput work.

Anthropic indicates that Fable 5.1 reduced the cost of typical workloads by about 25%, while compared to Opus 5, Opus 5.5 lowered the cost of individual tasks for customers by about 40%. These reductions reflect both price cuts and improvements in efficiency, rather than direct discounts on each task. Anthropic provides a detailed introduction to the lower caching read pricing of Fable in the model release notes.

This strategy has its commercial logic: cheaper and more powerful AI may encourage customers to increase their usage. Anthropic indicates that historically, users have tend to increase their consumption after such announcements.

However, the latest quarterly report from the company also demonstrates how difficult it is to synchronize demand with supply. Anthropic typically reserves capacity 12 to 18 months in advance of when the computing power becomes actually available, which means that a significant portion of these expenditures remains relatively fixed in the short term. This practice has also made it "famous" among AI enthusiasts, as quotas are often easily exhausted.

The quarter included a short-term committed computing power of $5 billion. Anthropic indicates that the cost of this portion of computing power is significantly higher than that of similar capacity locked in through multi-year contracts. This expenditure contributed to the company's shift from an operating profit of $974 million in the second quarter to an operating loss of $567 million in the third quarter.

The longer-term interests are even greater. The IPO documents of Anthropic indicate that the company is preparing to go public, which also brings it into the spotlight of Wall Street. Updated disclosures show that the total amount of its leasing and procurement commitments is approximately 895.4 billion US dollars, of which nearly all are related to computing and infrastructure.

Reuters previously reported that this document raised doubts among outsiders about whether infrastructure partners would be able to raise sufficient funding to deliver all of this production capacity. For now, Google and Nvidia seem to be the biggest beneficiaries of Anthropic's growth, as their core businesses rely on their infrastructure.

At present, the demand for Claude is still growing rapidly. The more difficult question to answer is whether Anthropic can reduce the costs of building and operating these models even faster while the models continue to see price reductions. Currently, OpenAI, which is also preparing for IPO, offers competitive models with more usage reset mechanisms and more lenient usage quotas.

This is precisely the main issue that prevents Anthropic from raising prices.