Morning Edition · Sunday, July 5, 2026

Anthropic Cuts Off Claude Code in China Over Distillation Claims, Alibaba Bans It Back

The dispute turns a developer tool into the newest front in the US-China split over frontier models.

Anthropic has begun blocking access to its coding tool, Claude Code, for developers based in China. The company accuses several Chinese laboratories of extracting Claude outputs to train competing models. According to AI ML Big Data, a Russian-language technical channel, Anthropic named DeepSeek, Alibaba, MiniMax and Moonshot AI, and said the firms were processing Claude responses at scale to copy its capabilities into their own systems.

Alibaba responded with its own restriction. According to AI Post, citing Reuters, the company will bar its employees from using Claude Code internally starting July 10, pointing to alleged backdoor and data-theft risks. AI Post reports that the ban follows an accusation from Anthropic two weeks earlier that Alibaba had extracted roughly 28.8 million Claude outputs.

Both figures and the copying allegations come from the two companies and secondhand channels rather than independent forensic review. Neither side has published logs or model-provenance evidence that a third party could reproduce. What is verifiable is the direction. A US laboratory is now restricting a developer product by nationality, and a major Chinese cloud vendor is treating that same product as a security threat.

Consider who benefits. Anthropic gains a legal and public-relations basis to defend its training investment and to justify the identity-based access controls it has been introducing anyway. Alibaba gains justification to standardize on domestic tools it already builds. Each move pushes developers on both sides toward the model stack aligned with their government.

Veracity: Corroborated

82/100

If true, who benefits: Both governments' decoupling agendas, plus each firm's domestic lock-in: Anthropic protects its training-investment moat and identity-based access controls, while Alibaba accelerates enterprise migration to its own Qwen and Qoder stack.
The nuance: The 28.8 million-exchange distillation figure rests on Anthropic's own account with no reproducible logs, while Alibaba's "backdoor" label describes real anti-abuse tracking code (time-zone, proxy and Chinese-firm keyword detection) that an Anthropic engineer acknowledged, not proven espionage.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

The decoupling has moved past chips and cloud contracts to the everyday tools engineers use to write software. If access to a coding tool is now assigned by nationality, teams operating across the US-China divide will need parallel toolchains, and the practical cost of that duplication falls on multinationals and on any firm serving both markets.

What to watch

Whether Anthropic publishes reproducible evidence of distillation, or the claim remains an assertion. That determines whether this is an enforceable intellectual property case or a market-access maneuver.
Whether other US laboratories such as OpenAI and Google follow with nationality-based restrictions on their own agent products, which would signal an industry norm rather than one company's stance.
How fast Chinese firms adopt domestic coding tools internally, which shows how deep the tooling split really runs.

Observations to monitor, not financial advice.

1 source

Source: Polylog editors

Part of a tracked trend

US-China Technology Decoupling

The US and China keep widening the separation of their technology sectors from chips toward platforms, models, and corporate access, and each new designation or countersuit pushes firms to pick a side and rebuild supply chains around the bloc they serve.