Ex-Anthropic researcher Coxon says Mythos could be retrained far more cheaply

Key points

• Coxon says he left Anthropic after about four months, citing a slide from its IPO roadshow on training compute.
• The chart's compressed axis hides large jumps, but Coxon reads the steep curve as accelerating algorithmic efficiency gains.
• He claims Mythos, a spring capability jump, fits this trend and could now be retrained with a fraction of the compute.
• Coxon wants frontier model development paused next year, arguing a pause would harm OpenAI and Anthropic.

Why it matters for marketers: Marketers covering Anthropic's IPO or model launches should note the compute-efficiency claim is one former staffer's opinion, unverified in the source.