Why Anthropic’s Gross Margin Cannot Be Directly Compared to Traditional Software Companies
Anthropic’s S-1 filing discloses that over 80% of its revenue in the third quarter of 2026 came from model versions released within the preceding six months. OnlyCFO points out that because model development costs are classified as research and development (R&D) rather than cost of goods sold (COGS), Anthropic’s reported gross margin is not directly comparable to that of traditional software companies; if 50% of R&D were reclassified into COGS, the gross margin for the first half of 2026 would drop from its apparent high level to approximately 55.5%.
Anthropic disclosed a set of easily overlooked figures in its S-1 filing: during the three months ended September 30, 2026, over 80% of its revenue was generated from model versions released within the previous six months. This implies that Anthropic’s current revenue is heavily dependent on its most recent models, while the development costs for these latest models are entirely expensed as research and development (R&D) and do not enter the cost of goods sold (COGS).
The logic for traditional software companies differs. A company develops a product once and can benefit from the same codebase for subsequent years; R&D is a one-time investment, while COGS primarily consists of hosting, support, and minor maintenance. Consequently, gross margin can truly reflect the economics of unit delivery. In Anthropic’s case, models have a shelf life; today’s revenue requires tomorrow’s R&D to sustain it. Here, R&D is not an investment in the future but a necessary expense to maintain current revenue.
OnlyCFO performed a sensitivity calculation: if 50% of R&D expenses were allocated to COGS, Anthropic’s gross margin for the first half of 2026 would be approximately 55.5%. While still respectable, the gap between this figure and the reported number is sufficient to demonstrate that the two metrics are not calculated on the same basis. The author also notes that if Anthropic achieves recursive model training in the future—requiring less R&D to produce new models—this issue might diminish, but this remains an unfulfilled condition.