Anthropic accuses Alibaba of using 25,000 fake accounts to harvest Claude data for training Qwen models

https://www.bloomberg.com/news/articles/2026-06-24/anthropic-accuses-alibaba-of-illicitly-accessing-its-ai-models

imagine ur at Anthropic, u get a new colleague from Alibaba and he's suspiciously, incredibly familiar with the Claude API from Day 1

“It not fair, I stole it first”—Dario

It is hilarious to see the large scale internet copyright scraping operation on OUR data by Anthropic, and they got sued and have to pay settlements on, if pointing to someone that may be scraping what Anthropic already stole.

@kimmonismus What a day, it's still going.

There will probably be an Anthropic blog post about this shortly.

@AndrewCurran_ crazy, isnt it

@kimmonismus 28.8 million Claude exchanges in 90 days through 25,000 paid accounts.

Alibaba didn't steal the model. they bought it one prompt at a time. whatever they paid for those tokens is Anthropic's R&D credit for Qwen's next jump.

@kimmonismus Micron also just killed it on earnings.

@kimmonismus whereas Anthropic trained its models on public license/royalty free data obviously.

@appsicle_ Someone needs to do the math. Curious as well

napkin math:

28.8M exchanges / 25K accounts = 1,152 per account in 90 days. roughly 13 per account per day, blends right into normal SWE workloads.

at an average agentic exchange (6-15K tokens) and Opus 4.8 rates, that's $3-12M Alibaba paid Anthropic between April and June.

frontier-class supervised distillation dataset for the price of a single seed round. Anthropic just funded Qwen's next jump with a 90-day pilot.

@limalemonnn we need someone to do the maths

@leaving_tech At least someone honored my joke :D

@kimmonismus Liars! This is a psyop to try to get the US government to ban Chinese A.I. in the United States.

@JR_Openheimer @kimmonismus yet they all still do it anyways

@JR_Openheimer @kimmonismus probably not that far, it was open for 72 hours exactly so 3 days

@kimmonismus 25,000 accounts using Claude to orchestrate Qwen to reproduce certain behaviour? Distillation could actually be acceptable normal use, but anthropic loves to co talk about it.

Qwen should continue.

@kimmonismus 28.8 million exchanges is a lot of data, but distillation isn't magic. You still need solid architecture and engineering to make it useful. Raw output volume doesn't equal model quality.