Jacob Coxon says Anthropic IPO chart shows accelerating compute efficiency

Research experience, and a disagreement over progress

Coxon worked at OpenAI from 2023 until May 2026, then joined Anthropic and resigned on September 8 after roughly four months, according to the account of his departure. His relevant authority comes from working on model training at both developers. He is now arguing publicly for a pause, rather than speaking on Anthropic’s behalf.

The account says he forfeited unvested equity by leaving and had raised concerns about rapid progress toward self-improving AI before his latest comments. That gives the dispute a personal financial dimension, although forfeited equity does not establish that his technical interpretation is correct.

The account describes an offering of approximately $64 billion, but that is not all money for Anthropic. Its November 10 IPO prospectus proposes about 801.4 million company shares and 23.6 million shares from existing holders, priced at $70–$78 each. At the upper price, multiplication puts the combined offering at $64.35 billion, of which about $62.5 billion would be gross proceeds for Anthropic, before fees and excluding any additional underwriter purchases. The company would receive none of the selling holders’ proceeds.

For prospective investors, efficiency could mean reaching an existing performance target with fewer resources or using the same resources to pursue a more capable model. Those are different economic choices. A pretraining-efficiency curve alone establishes neither lower total spending nor cheaper service for customers.

Coxon wants frontier model development paused in 2027 so competitors can catch up. In his November 13 post, he argues that this would make human-level intelligence a commodity and would hurt OpenAI and Anthropic. Those are his predictions about competition and capabilities. His proposal depends on rivals being able to reproduce the capabilities and on the scope of the pause; the post does not set out an enforcement mechanism or show that competitors have the necessary computing resources.

Ho and colleagues estimated that compute requirements at fixed benchmark performance halved about every eight months. This historical finding does not independently verify Anthropic’s current trajectory. Source: Algorithmic progress in language models.

Efficiency gains have a measurable precedent

A 2024 study by Anson Ho and colleagues examined more than 200 language-model evaluations on Wikitext and Penn Treebank spanning 2012–2023. It estimated that computation required to reach a fixed performance threshold halved approximately every eight months, with a 95% confidence interval of roughly five to 14 months.

The authors also found that increased compute contributed more to overall performance improvement during that period than algorithmic progress did. They cautioned that the benchmark data were noisy.

That provides precedent for substantial efficiency improvements alongside expanding compute use. It is not a like-for-like test of Anthropic’s slide: the study covers different historical models and benchmarks, and does not verify accelerating gains at one company or an imminent transition to automated AI research.

What would establish the chart’s significance

The key missing evidence is the slide’s measurement method: the performance target held constant, the models and dates compared, the computation counted, and the underlying data. Those details would allow scrutiny of both the size of the efficiency gains and whether their rate is accelerating.

Separate evidence would be needed to connect those gains to reliable automation of algorithmic research. For investors, the unresolved question is how much those gains change the resources needed for future models. For Coxon’s pause proposal, it is whether competitors could catch up under a workable set of rules—and whether that would deliver the safety and competitive benefits he predicts. The reported chart does not settle either question.