Anthropic has given investors its first concrete estimate of how much money “distillation” by competing AI labs was actually worth to its bottom line.
According to a Financial Times report, Anthropic told investors on Wednesday that distillation accounted for 5 to 7 percent of its revenue, and that it had cut this back “materially” in the third quarter of the year, according to four people with direct knowledge of the matter. The FT report notes that the startup has clamped down hard on the practice, where rival labs train their own models on the outputs of Anthropic’s more advanced ones, and that when competitors distil from Anthropic’s models, it shows up as heavy usage and spending on Claude.
That 5-7 percent figure is interesting because it’s the clearest sign yet that distillation wasn’t just a security or IP concern for Anthropic, it was inflating the company’s own revenue numbers, the very numbers investors are leaning on ahead of a planned IPO that could value the company at $2 trillion or more.
Why Anthropic went after distillers
Anthropic has spent much of this year calling out rival labs over the practice. In February, the company accused three Chinese AI labs, DeepSeek, Moonshot AI and MiniMax, of running what it called “industrial-scale” distillation campaigns, generating over 16 million exchanges with Claude through roughly 24,000 fraudulent accounts, in violation of its terms of service and regional access restrictions, since Anthropic doesn’t offer commercial access to Claude in China. MiniMax alone was blamed for more than 13 million of those exchanges, centered on agentic coding and tool orchestration, and was said to have redirected nearly half its traffic to a newly released Claude model within 24 hours of its launch.
Anthropic returned to the subject earlier this month. In an October 12 research post, the company said a model distilled from Claude’s full reasoning transcripts scored 80.4% on SWE-bench Verified, versus 71.8% when distilled only from summarized reasoning, which is essentially why reasoning traces, not just final answers, had become the real prize for distillers. The company said a bundle of defenses rolled into Claude Fable 5.1 and Opus 5.5, including summarizing rather than exposing full reasoning, stricter rate limits for repeat offenders and faster account bans, had cut what distillers could harvest by roughly 90-99%. That post named Zhipu’s GLM-5.3 and Moonshot’s Kimi as beneficiaries of the extracted data, and said tens of thousands of accounts tied to distillation campaigns had been banned, often within two hours of a campaign being detected.
A trade-off Anthropic was willing to make
The investor disclosure effectively confirms that shutting distillers out meant giving up real revenue, exactly the kind of trade-off Anthropic has openly defended. Anthropic isn’t new to tightening access for anyone it feels is using Claude in ways it didn’t sign up for. The company had earlier made it clear that Claude subscriptions would no longer cover usage on third-party tools like OpenClaw, pushing that kind of heavy usage onto separate paid bundles or API billing instead.
The timing of the disclosure is notable too. Anthropic is reportedly pushing to file for an IPO that could value it at $2 trillion or more, with backers expecting annualized revenue to hit $100-120 billion by the end of this year. Investors weighing that valuation will want to know how much of the company’s blistering growth is organic enterprise demand, the kind Anthropic has been chasing hard with pushes like its recent Claude Financial Services launch, versus inflated usage from rivals quietly mining Claude for training data. By putting a number on distillation and showing it’s already shrinking, Anthropic appears to be trying to get ahead of that exact question before it becomes a prospectus problem.