Coxon Says Anthropic Chart May Show Faster AI Training Efficiency Gains

Jacob Coxon, a former pretraining researcher at OpenAI and Anthropic, says a chart in Anthropic’s IPO roadshow may point to faster gains in training efficiency, a concern he has tied to his resignation and call for a pause in frontier AI development.

The chart compares the computing power needed to train models with comparable coding ability as algorithms improve. It shows that less compute may be needed for a similar coding result. Coxon focused on the curve’s increasingly steep slope despite its logarithmic vertical scale. He interprets that shape as a possible sign that the rate of efficiency improvement is accelerating, if the data are accurate.

The graphic does not quantify the trend. It contains three data points, and their plotted positions are marked as illustrative. It gives no compute multiples, so readers cannot calculate the size of the reduction or how quickly it is changing. The comparison concerns coding ability; it does not quantify other model capabilities.

The filed roadshow materials were dated Nov. 10, about two months after Coxon announced his resignation on Sept. 8. The filing date does not establish when the slide was first created. Coxon has described the chart as one reason he left in September 2026.

Anthropic co-founder Tom Brown discussed the possibility that future versions of Claude could find ways to make training more efficient. Such methods could help produce stronger models, which could then identify further improvements. That potential cycle is a possibility described in the roadshow discussion; the chart does not quantify its effects.

Coxon has called for a pause in frontier AI development in 2027, arguing that it would give competitors time to catch up. He has said that broader access to human-level AI capabilities could weaken the advantages that companies build around leading models.