White House Accuses Moonshot AI of 'Large-Scale' Theft From Anthropic's Fable 5
The White House has made its sharpest accusation yet against a Chinese AI lab. Michael Kratsios, head of the White House Office of Science and Technology Policy, stated on the record that Moonshot AI used distillation from Anthropic’s Fable 5 to build Kimi K3 — the 2.8 trillion-parameter open-weight model that, when released on July 17, posted results within 2 percentage points of Fable 5 on SWE-bench Verified.
“We have information that Moonshot distilled Fable for the development of K3,” Kratsios said.
The accusation goes beyond prior distillation claims. Anthropic named Alibaba in June, documenting 28.8 million Claude API exchanges from 25,000 fake accounts used to extract training signal. The Moonshot accusation is the first to come directly from the White House, and comes alongside a separate allegation: Moonshot acquired advanced Nvidia AI chips through restricted channels.
Kimi K3’s Numbers in Context
The model at the centre of the accusation: Kimi K3, a 2.8T MoE open-weight model from Beijing-based Moonshot AI, released July 17 at $3/$15 per million tokens. Its benchmark performance:
| Benchmark | Kimi K3 | Fable 5 | Gap |
|---|---|---|---|
| SWE-bench Verified | 93.4% | 95.0% | 1.6 pts |
| LiveBench Overall | 78.5 | 80.8 | 2.3 pts |
| Harvey LAB-AA Legal | 26.7% | ~14% | +12.7 pts |
| LiveBench Agentic Coding | 57.6 | 46.9 | +10.7 pts |
On pure SWE-bench Verified, Kimi K3 sits third globally — behind Mythos 5 (95.5%) and Fable 5 (95.0%). A 1.6 percentage point gap is smaller than the variance between evaluation runs. Moonshot maintains the model was developed independently.
Why Distillation Is Hard to Prove
Distillation accusations require differentiating between a model that trained on outputs from a target model versus one that simply achieved similar performance through independent means. The legal and technical bar is high.
In the Alibaba case, Anthropic had explicit API log evidence: 28.8 million exchanges, 25,000 accounts, coordinated query patterns. The Moonshot accusation is, at least in the public record, an intelligence claim without the same documentary trail. CyberScoop and the New York Post reported the statement but Kratsios has not released underlying evidence.
Moonshot has not issued a direct denial as of publication.
The Chip Allegation
The second accusation — restricted Nvidia chip acquisition — connects to a separate US enforcement concern. Export controls on advanced AI chips introduced under BIS regulations in late 2023 and tightened through 2025 are supposed to prevent Chinese labs from accessing chips above a certain performance threshold without licences. Kratsios said Moonshot obtained hardware in violation of those controls.
If accurate, this would represent both IP theft and export control circumvention in the same model development cycle.
Pattern and Stakes
This is the second time in two months the US government or a frontier lab has publicly named a Chinese AI company for distillation. The pattern matters: each major open-weight Chinese release since DeepSeek R1 has been followed by accusations or evidence of data extraction from closed Western models.
The immediate practical consequence: calls in Congress and within the White House for formal entity listing of Moonshot AI, which would restrict US cloud providers and AI companies from continuing to serve the model. Kimi K3 is currently available on multiple Western cloud inference platforms.