Elon Musk's xAI spent months using outputs from Anthropic's Claude chatbot to help train its coding models, according to a report from The Information. The account suggests the company continued the practice even after Anthropic cut off official access earlier this year.
The report says xAI directly trained its coding system on Claude-generated responses. When Anthropic revoked access in January, engineers at xAI allegedly kept collecting Claude output through personal accounts and by using the intermediary service Blackbox AI. The reported arrangement highlights the competitive pressure among frontier AI companies as they race to improve coding performance and general model capability.
Training on another company's model outputs is a form of distillation, a technique where one system learns from the behavior or responses of another. The practice is common enough in the AI industry that Musk has previously defended xAI's use of outside models in court, saying the company had only "partially" used OpenAI models to help train Grok and describing the method as standard industry practice.
The report also portrays a company under internal stress. According to The Information, xAI's pretraining team has shrunk to fewer than five people. It says four leads focused on Grok's coding work left the company within a few months, alongside several co-founders.
The publication also reported that one employee accidentally deleted important training data, setting back work by two to three weeks. Taken together, the departures and data loss suggest that xAI has faced operational challenges while trying to keep pace with rivals in the fast-moving AI model market.
The company has also had to make use of the large amount of computing capacity Elon Musk has assembled over the past few years. The report notes that some of that infrastructure is now being rented out rather than used solely for xAI's own training efforts. SpaceX is said to be leasing compute to Anthropic, while Google is also reported to be a customer.
That shift is notable because Musk has spent heavily on hardware to support xAI's ambitions, including efforts to build and expand massive GPU clusters. The report suggests some of that capacity is not currently being dedicated to xAI model training in the way Musk may have intended.
Anthropic and xAI have not publicly detailed the scope of the alleged training arrangement in the source material. The new report is likely to intensify scrutiny over how AI companies gather training data, especially when they depend on the outputs of rival systems. It also raises questions about whether access controls are enough to prevent model builders from continuing to draw on another company's services through alternate channels.
For now, the account adds another layer to the broader debate over data sourcing, model distillation, and the increasingly aggressive tactics AI developers use to gain an edge in coding and other competitive benchmarks.