Former OpenAI and Anthropic researcher Jacob Coxon says Anthropic’s own investor materials point to accelerating improvements in model-training efficiency. He highlighted a chart from the company’s IPO roadshow that tracks how much capability Anthropic obtains from a given amount of compute over time.
Coxon argues that the chart’s steep rise remains striking even though its axis compresses tenfold and hundredfold changes. He interprets it as evidence that algorithmic progress is accelerating rather than moving at a constant rate. In his view, Anthropic could already retrain its Mythos model using only a fraction of the computing power required for the original run.
That conclusion is Coxon’s interpretation of the chart, not a disclosed retraining result from Anthropic. The report does not provide a specific compute ratio, independent reproduction or enough underlying data to establish the exact rate of improvement.
Coxon says the prospect of automating further algorithmic advances contributed to his decision to leave and supports his call for a pause in frontier development. He argues that a pause would allow competitors to catch up and reduce concentration around Anthropic and OpenAI. The immediate significance is therefore less a verified performance milestone than a public warning from a former insider about how he reads the company’s efficiency trajectory.