Former researcher Coxon says Anthropic could retrain Mythos with far less compute

Why a researcher left Anthropic

Jacob Coxon, a former researcher at OpenAI and Anthropic, has explained why he left the company after roughly four months. According to The Decoder, he points to a slide from Anthropic's IPO roadshow as a key reason.

The slide charts how much compute, meaning processing power, Anthropic needs over time to train its models. Coxon says the curve still climbs steeply even though its axis is compressed, which he reads as evidence that algorithmic efficiency is improving at an accelerating rate.

What the slide suggests about Mythos

Mythos is Anthropic's model that delivered a major capability jump in the spring. Coxon argues it was not a one-off result but one point on a steep, accelerating trend.

He claims Anthropic could already retrain Mythos using a fraction of the compute originally needed. These are his own statements, reported by The Decoder, and Anthropic has not confirmed them.

A call to pause frontier models

Coxon continues to call for a pause on frontier model development next year so competitors can catch up. He argues such a pause would turn human-level AI into a commodity and would mainly harm OpenAI and Anthropic.

He also says algorithmic improvements are close to being automated, a concern he links to wider fears that self-improving AI systems could become difficult to control. These views are contested and remain his personal position.

Why it matters

If Coxon is right, the cost of reaching a given level of AI capability is falling quickly, which could reshape competition among frontier labs. The claim rests on a single public slide and his interpretation of it, so readers should treat it as unconfirmed.

The episode also shows how internal company material, released for investors, can become evidence in public debates about AI safety and pacing.