Anthropic Discloses 5-7% Revenue from Model Distillation, Actively Cutting It to Block Rivals
Anthropic confirmed in an investor update that revenue generated from competitors using its API for model distillation accounts for 5-7% of total sales, a segment significantly reduced in Q3. While this strategic cut impacts short-term financial growth, it aims to prevent labs like DeepSeek and Moonshot from cheaply replicating Claude's capabilities by blocking access to high-quality training data, highlighting the tension between monetization and defensive technical moats.
Full Report
Anthropic informed investors on Wednesday that revenue from "distillation" accounts for 5% to 7% of its total revenue, and the company has already significantly reduced this portion in the third quarter. The news comes from the Financial Times (FT), citing four people directly familiar with the matter; The Information reported the same range.
What is distillation? It involves using responses from more powerful models as training data to train one's own model. Competing labs can query Claude extensively, collect the responses, and feed them into their own models, thereby acquiring some of Claude’s capabilities at a lower cost.
Since it is essentially "learning," why does it count toward Anthropic’s revenue? Because these queries go through paid interfaces, where each call is billed based on usage. Distillation often requires millions or even tens of millions of queries. According to the FT, distillation "involves significant usage and spending." In other words, this revenue represents money actually paid by competitors, albeit for the purpose of training rival AI models. Anthropic’s commercial terms of service explicitly prohibit using its services to "train competing AI models."
How much is 5% to 7%? The reports did not specify which period this percentage applies to, so we can only estimate. A calculation circulating online uses Q2 revenue of $11.5 billion, resulting in an estimated $575 million to $800 million. Note that this is an inference, not a figure stated in the reports. The $11.5 billion number comes from company documents seen by Bloomberg in August, described as "over $11.5 billion" in preliminary figures, which Anthropic has not publicly confirmed.
The reports also did not identify which companies contributed this revenue. In February of this year, Anthropic published a post naming DeepSeek, Moonshot AI, and MiniMax, stating that the three firms used approximately 24,000 fake accounts to bypass restrictions on providing Claude services to China, engaging in over 16 million conversations with Claude for distillation purposes.
Actively cutting this revenue source comes at a cost. Anthropic is preparing for an IPO. The same briefing revealed that Q3 revenue was $16.8 billion, up about 46% from Q2, while Q2 had more than doubled compared to Q1, indicating slowing growth rates. According to reports, the company told investors that the reduction in distillation revenue would drag down this quarter’s numbers; the trade-off is making it harder for competitors to catch up by copying Claude’s answers.
↳ Quote @simonw (Simon Willison): "Anthropic told investors on Wednesday distillation accounted for 5 per cent to 7 per cent of revenue"
If true, and Q2 revenue was $11.5B, then that's more than half a billion dollars spent by other labs to distill Anthropic's models! https://www.ft.com/content/11632d03-358c-454b-af95-d06d4072932b?syn-25a6b1a6=1