← The feed Many Minded · The brief
GOODS · forward · impact 2/5 · 2026-11-14 · Anthropic

Anthropic's retraining claim suggests path to cheaper AI models

Former Anthropic researcher Jacob Coxon claims the company can retrain its Mythos AI model using a fraction of the compute originally needed for training. This assertion draws from a logarithmic slide

Former Anthropic researcher Jacob Coxon claims the company can retrain its Mythos AI model using a fraction of the compute originally needed for training. This assertion draws from a logarithmic slide in Anthropic’s IPO roadshow showing projected 'pre-training compute multipliers' from July 2024 to October 2026. Anthropic states the data represents internal estimates for illustrative purposes only. If validated, such efficiency could lower AI training costs, accelerating accessible AI solutions. However, Coxon’s claims rely on his personal assessment of algorithmic progress, and the timeline is a forward-looking projection. The next step is whether Anthropic will pause frontier model development as Coxon advocates to enable broader AI commoditization. This remains unverified but hints at potential pathways for reduced AI training expenses if the efficiency gains materialize in practice.

Source: The Decoder