Alibaba Accused of Distilling Claude, Then Went Closed-Source: Don’t Just Watch the “Stealing” Drama — What’s Really Changing for Open-Source AI

 

First, the three numbers: 25,000 fake accounts, 28.8 million conversations, over 44 straight days. According to Anthropic, these accounts were systematically extracting Claude’s core strengths — especially coding and agent reasoning. This isn’t casual scraping. It’s large-scale distillation of a frontier model.

A lot of people’s first reaction is “China is stealing American AI again.” Distillation happens in the industry, but today we’re not debating that. Instead, let’s focus on three more important things.

1. Anthropic’s Letter: The Scale and Timeline

On June 24, 2026, Bloomberg published a letter Anthropic sent to the U.S. Senate Banking Committee and the White House, dated June 10 — right before an AI hearing.

The letter claims that between April 22 and June 5, roughly 25,000 fake accounts had 28.8 million conversations with Claude, specifically targeting code generation and agent reasoning. Back in February, Anthropic had already complained about DeepSeek, Moonshot, and MiniMax combined — 24,000 accounts and 16 million queries. This time, Alibaba alone exceeded the total of the previous three.

Alibaba has not publicly responded so far.

2. Alibaba Is Changing Direction: From Open-Source to Closed-Source

What’s more noteworthy is what was happening inside Alibaba.

On March 4, Lin Junyang, the key technical leader behind Qwen, posted on X that he was “retreating” and leaving the team he loved. He later left to start his own company. Around the same time, the post-training lead and the Qwen Code lead also departed.

April 22 is exactly when the alleged large-scale distillation began. On May 19, Qwen 3.7-Max launched — but it was no longer open-source (Qwen 3.6 had been open).

The timeline is tight: the main open-source advocate left, and about seven weeks later the distillation started. The closed-source flagship dropped right in the middle of that period. A new team with less historical baggage chose the fastest shortcut — distill first, then ship closed-source.

3. Anthropic Is Also Trying to Save Itself

Anthropic’s own situation isn’t comfortable either.

On June 9 they launched two flagship models, Fable 5 and Mythos 5. Three days later, the U.S. Commerce Secretary sent a letter ordering them taken down globally — including inside the United States — because their safety guardrails had been bypassed. Ironically, Anthropic’s own letter warned that distilled models often lack proper safety protections. Their own models got hit first and remain restricted.

Anthropic is also preparing for an IPO at a reported $965 billion valuation. They need their strongest models available to stay ahead. Meanwhile, OpenRouter’s Fusion and Japan’s Sakana AI Fugu have already reached similar performance through multi-model orchestration. If their top models stay blocked for long, the lead could slip.

It’s worth noting the timing: Alibaba was added to a U.S. defense-related list on June 8. Anthropic sent its letter on June 10. Alibaba later sued the Pentagon to be removed, and China responded with countermeasures against U.S. defense firms. After Bloomberg published the letter, Alibaba’s stock dropped about 3% in the short term.

4. The Real Issue: The Full-Spectrum Open-Source “Shelf” Is Disappearing

Whether the distillation was right or wrong, the user agreement is clear — if you use the product, you follow the rules. But the bigger, quieter change is what happens after Alibaba moves to closed-source.

Plenty of Chinese labs now open-source huge models: Kimi (~1T parameters), MiniMax (~229B), Zhipu GLM, Xiaomi MiMo, etc. These are impressive, but most individuals and small teams can’t actually run them. They’re built for large cloud deployments.

What regular developers and researchers really need is a full ladder of open models — 0.6B, 1.7B, 7B, 14B, 32B, 72B and so on — so they can fine-tune and do post-training at the size that fits their project.

Meta’s Llama used to provide this ladder. Over time they reduced smaller open releases and moved toward closed models. Qwen had become the last major player still maintaining the complete spectrum — not just one giant flagship, but models across sizes that normal people could actually use and adapt. Sakana AI’s Fugu, for example, was built on Qwen 2.5 7B.

From Qwen 3.7 onward, that full ladder is gone. If you want to build something small, run on edge devices, or do custom fine-tuning without massive compute, the accessible open-source foundation is shrinking.

5. Three Final Points

• Don’t rush to pick sides. 

Both stories can be true at the same time: the U.S. using national security tools to shape the playing field, and Alibaba’s new team choosing closed-source plus fast-following as a business strategy. The reporter has self-interest; the reported party also took controversial actions.

• Technology has no borders, but business and politics do. 

At this stage of AI, the game isn’t just about who is smarter — it’s about who can enclose that intelligence inside their own APIs, blacklists, or IPO filings.

• Don’t let the “stealing” debate distract from the real question. 

After the last major player maintaining a full range of accessible open-source models steps away, can ordinary developers who can’t afford thousands of GPUs still build on top of frontier progress? That light going out affects far more people than just the big labs.

6. Two Questions Worth Discussing

• Was Alibaba’s shift to closed-source a pragmatic business move, or did it abandon its earlier open-source vision?
• If even Alibaba is pulling back, do you still believe open-source AI can remain genuinely accessible to regular developers and smaller teams?

Feel free to share your thoughts in the comments.