Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon

Jacob Coxon, a former researcher at both OpenAI and Anthropic, says a slide from Anthropic's IPO roadshow was a major factor in his decision to leave the company after roughly four months. The chart plots the compute Anthropic needs over time to train its models, with a logarithmic y-axis and a compressed x-axis so that tenfold and hundredfold increases are visually subdued. Even so, the curve climbs steeply, and Coxon reads that as evidence that algorithmic efficiency is improving at an accelerating rate. He calls it an insane plot and says he is surprised the company made it public.
Coxon ties Anthropic's Mythos model to the same trend. Mythos, which he describes as a major capability jump in the spring, was not a one-time deviation in his view but a point on a superexponential curve. He states that the company could already retrain Mythos with a fraction of the compute originally required. The implication is that the cost of reaching a given capability level is falling fast, not just steadily. A logarithmic axis compresses large multipliers, so a curve that still rises visibly on such a plot represents growth that is large in absolute terms. That is the basis for Coxon's claim: even with the axis flattened, the slope remains steep enough to suggest compounding gains. A likely trade-off is that internal estimates and illustrative data, which Anthropic flags on the slide, can overstate precision, and the underlying measurements are not public.
Coxon also argues that algorithmic improvements are on the verge of being automated. That claim underpins his call to pause frontier model development next year so competitors can catch up. In his framing, a pause would commoditize human-level intelligence and would harm only OpenAI and Anthropic. His position follows earlier public warnings in September about the risks of advanced AI systems, warnings that other industry figures echoed. One concern he and others have raised is that advanced, self-improving AI could get out of control. The new comments connect that broad safety argument to a specific piece of evidence about falling training costs.
For context, training compute has been a central lever in frontier AI. If the compute needed to reproduce a model's capability drops sharply, the barrier to matching that capability falls with it, which changes how defensible a lead based on scale alone can be. This suggests the strategic value of raw compute advantages may erode faster than hardware procurement cycles assume. It also suggests that smaller labs and open efforts could close gaps sooner than raw spending figures imply, though that is inference from Coxon's characterization rather than a demonstrated result.
Why it matters: Developers and product teams should expect model capability per dollar to keep improving quickly if Coxon's reading holds, which shortens the useful life of assumptions baked into cost and capacity planning. For labs, the competitive moat may shift from scale toward algorithmic efficiency and the ability to automate research itself. The call for a pause is a policy proposal, not a decision, and whether competitors would follow remains an open question.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Former OpenAI and Anthropic researcher Jacob Coxon points to a slide from Anthropic's IPO roadshow as a reason he left, saying it shows the company's algorithms are improving not just quickly but at an accelerating rate. Coxon continues to call for a pause on frontier model development so competitors can catch up and human-level AI can become a commodity. The article Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon appeared first on The Decoder .