Anthropic R&D spending surged 5.3x as model progress nearly doubled

The scale of Anthropic's spending

Anthropic's recent Form S-1 filing reveals a massive increase in research and development costs. The company spent $1.85 billion on R&D during the first half of 2025, but that figure jumped to $9.74 billion in the first half of 2026.

This represents a 5.3x increase in spending over a single year. The filing specifies that these R&D costs primarily cover compute and infrastructure for model development, internal usage, and personnel costs.

As reported by FourWeekMBA, this surge in spending provides a clear look at the capital requirements currently needed to push the frontier of artificial intelligence. It shows a move toward much larger infrastructure commitments to achieve the next generation of models.

Measuring model progress

To understand if this spending is yielding results, researchers at METR analyzed Anthropic's internal data. They used a metric called the Epoch Capabilities Index (AECI), which is Anthropic's version of a benchmark used to measure model capabilities.

METR found that the rate at which Anthropic's models improve on this index accelerated about 1.9 times since October 2025. Before that date, the trend showed an improvement of about 15 points per year, which then steepened to about 29 points per year.

While the models are indeed improving faster than they were previously, the speed of that improvement does not keep pace with the massive spike in spending. The data suggests that while the

The mechanics of AI capital

The financial landscape of AI development is shaped by the relationship between capital expenditure and depreciation. A core concept in this area is how the lifetime of hardware chips drives the overall profits of an AI company. Understanding these dynamics is essential for evaluating how companies manage their massive infrastructure investments.

The Business Engineer framework explains these concepts through an evergreen explainer on AI capex. This perspective helps clarify how the physical assets used to train models impact the long-term financial health of the organizations building them. It provides a foundation for understanding why the costs reported in filings are so significant to the industry's bottom line.

Detailed operating expenses

Anthropic's Form S-1, which was filed on October 26, provides a detailed table of operating expenses categorized by function. This table allows for a granular look at how the company allocates its resources across different departments and operational needs.

By examining these specific categories, analysts can see how research and development is separated from other costs. The filing ensures that the reported spending on compute, infrastructure, and personnel is clearly defined within the company's broader financial structure. This transparency is vital for assessing the scale of the company's operational growth.

The AI Supercycle framework

This story aligns with the Business Engineer framework regarding the AI Supercycle. This framework suggests that the current era of AI is defined by capital commitments that exceed the unit economics of any single application while simultaneously producing real productivity gains. It acknowledges that these two conditions can exist at the same time.

The data from Anthropic and METR illustrates this tension perfectly, as spending rises faster than the measured rate of progress. This framework is used by over 95,000 executives who read Business Engineer to understand the AI strategy frameworks used by major models. It helps leaders navigate the reality of high costs and high rewards in the current market.

Interpreting the data

It is important to note that the relationship between R&D spending and model progress is descriptive rather than causal. Neither the Anthropic filing nor the METR assessment establishes a direct causal link between the amount of money spent and the specific points gained on a benchmark.

The ratio of spending to progress is a way to observe trends, but it does not account for every factor involved in model development. Because R&D spending also pays for work that does not appear on the Epoch Capabilities Index, the per-dollar calculation is a simplified view of a much more complex operation.