Ex-Anthropic Researcher Says Retraining Mythos Needs Far Less Compute

Former Anthropic researcher Jacob Coxon has claimed that retraining the company's Mythos model now requires only a fraction of the original compute. If true, the assertion implies that algorithmic efficiency gains could dramatically lower the cost of training frontier models, challenging the industry's compute-scaling orthodoxy.

Coxon cited a chart from Anthropic's IPO roadshow—showing steep growth in training compute demand—as a key reason for his departure, saying it also revealed that algorithmic efficiency is improving at a super-exponential rate. He advocates a pause on frontier model development next year to let competitors catch up, making human-level intelligence a widely available commodity. However, these claims are personal remarks without official Anthropic data or papers, and their accuracy remains unverified.

If confirmed, the trend could reshape the AI landscape: lower training barriers may attract more players and intensify model commoditization, while leading firms may need to reassess the returns on massive compute investments. Safety and regulatory debates would also grow more complex as capabilities diffuse.