Ex-Anthropic researcher Coxon: Mythos could be retrained with a fraction of the original compute

Jacob Coxon, a former OpenAI and Anthropic researcher, says Anthropic could already retrain its Mythos model with "a fraction of the compute originally needed." Coxon made the claim on X, according to The Decoder, and tied it to a slide from Anthropic's IPO roadshow.

The slide

Coxon says the slide was a key reason he left Anthropic after only about four months. It charts how much compute Anthropic needs over time to train its models. The image caption identifies the metric as the "Pre-training compute multiplier" for code, plotted on a logarithmic y-axis, with data running from July 2024 to October 2026.

Because the axis is compressed, a 10x and a 100x increase barely look different. The curve still climbs steeply. Coxon says this shows algorithmic efficiency is improving "not just quickly but at an accelerating rate." "It's an insane plot and I'm surprised they made it public," he writes.

Anthropic states on the slide that the data points are based on internal estimates and are for illustration only.

Coxon's claims about Mythos

Coxon describes Mythos as a major capability jump in the spring. He argues it was part of the same trend rather than an outlier: "Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed."

These are Coxon's interpretations of the chart. Anthropic has not confirmed them in the material cited.

Call for a pause

Coxon says algorithmic improvements are "on the verge of being automated." On that basis he wants frontier model development paused next year so competitors can catch up. A pause, he writes, would "commoditize human-level intelligence. It would only harm OpenAI and Anthropic."

Coxon previously trained AI models at OpenAI and, most recently, at Anthropic. He drew wide attention in September with public warnings about the risks of advanced AI systems, which other figures in the industry echoed. One concern behind those warnings is that advanced, self-improving AI could get out of control.

What this means

The slide is Anthropic's own material, but the reading of it is Coxon's. The company labels the figures internal estimates meant for illustration, so the chart does not establish the size of any efficiency gain. Coxon's claim that Mythos could be retrained with a fraction of the original compute is an inference from the curve, not a disclosed figure.

Coxon's argument links two points: algorithmic gains that he says are accelerating, and a pause that he says would commoditize human-level intelligence. His case depends on the chart's trend holding and on algorithmic research being close to automated. Both are his assertions, not confirmed facts.

Coxon also says he left Anthropic over this slide, which puts an internal view of the company's efficiency trajectory into public debate.