xAI reportedly trained its coding models distilling Claude

Elon Musk’s artificial intelligence venture, xAI, has been exposed for engaging in a prolonged, covert “cat-and-mouse” game with competitor Anthropic. According to an investigative report by The Information, xAI engineers spent months utilizing a technique called knowledge distillation—directly training Grok’s upcoming coding models on the outputs generated by Anthropic’s flagship model, Claude.

The revelation underscores the fierce, often morally grey battle for high-quality synthetic and reasoning data as frontier AI labs rush to build advanced software engineering agents.

1. The Timeline of the Extraction Campaign

Knowledge distillation is a common machine learning technique where a developer prompts a highly capable “teacher” model and uses its high-quality responses to train a smaller or specialized “student” model. While tech companies routinely distill their own flagship models into lighter open-weight versions, using a direct competitor’s model to train a rival platform typically violates standard corporate terms of service.

The data-harvesting operation by xAI reportedly progressed through three distinct phases as Anthropic repeatedly tightened its security perimeters:

Phase 1 (Official Access): xAI engineers initially scraped data through authorized enterprise API keys. This avenue abruptly closed in January 2026 when xAI co-founder Tony Wu informed employees that Anthropic had officially identified and revoked their access channels.
Phase 2 (Personal Account Scraping): Refusing to halt the project, xAI developers pivoted to scraping data through an array of personal subscriber accounts. Anthropic caught on to the subtle traffic signatures and executed targeted bans on the associated accounts.
Phase 3 (Intermediary Exploitation): In a final effort to obfuscate their origin trail, xAI routed their automated queries through Blackbox AI, a popular developer platform acting as an encrypted intermediary network. This proxy method allowed xAI to continue benchmarking and distilling Claude data until mid-May 2026.

2. A Chaotic Backdrop at xAI

The reliance on competitor data highlights significant, growing pains behind the scenes at Musk’s startup. Despite commanding an astronomical amount of hardware compute, xAI’s internal engineering pipeline has reportedly been plagued by staffing crunches and operational mishaps:

Pretraining Team Atrophy: The core pretraining team at xAI has shriveled to fewer than five people following a wave of sudden departures.
Leadership Drain: Within just a few months, four separate Grok code engineering leads walked out the door, alongside several foundational co-founders.
The Data Deletion Blunder: Compounding the talent drain, an engineer accidentally deleted an entire batch of critical training data from xAI’s clusters, effectively erasing two to three weeks of non-stop work.

3. “An Industry Norm”

Elon Musk has subtly downplayed the controversy. During his ongoing legal battle with OpenAI in May 2026, Musk openly admitted under oath that early versions of Grok relied “partially” on OpenAI’s models for baseline guidance, dismissing cross-model training as an open “industry norm”.

The practice is undeniably widespread. Anthropic recently published a security breakdown exposing massive, systematic distillation campaigns targeting Claude from prominent Chinese AI laboratories—including DeepSeek, Moonshot AI, and MiniMax—which collectively deployed over 24,000 fraudulent accounts to harvest 16 million reasoning exchanges.

4. The Grand Irony: From Competitors to Landlords

The final, highly unusual twist in the Anthropic-xAI dynamic occurred in mid-May 2026. Facing a catastrophic computational bottleneck for its own upcoming models, Anthropic signed a massive deal to lease data center compute power directly from xAI to the tune of $1.25 billion per month.

┌────────────────────────────────────────────────────────┐
│  xAI extracts Claude's coding data to build Grok       │
└───────────────────────────┬────────────────────────────┘
                            │ (Cat-and-Mouse Game)
                            ▼
┌────────────────────────────────────────────────────────┐
│  Anthropic pays xAI $1.25B/month to use its servers     │
└────────────────────────────────────────────────────────┘

While Musk’s engineers were leveraging Claude’s logic to patch Grok’s gaps, Anthropic’s multi-billion dollar cash injections were actively cushioning xAI’s steep operational losses, helping secure a critical growth narrative for SpaceX’s upcoming Wall Street IPO.

To understand the broader technical and security implications of how frontier labs protect their models from being copied by rivals, you can watch Caught Distilling from Claude? This video breaks down Anthropic’s official technical disclosures regarding how they track, trace, and attempt to shut down coordinated data extraction campaigns.

Search for an article