Ex-Anthropic researcher Jacob Coxon says a roadshow chart of Anthropic's compute multipliers helped drive him out
What's new? Jacob CoxonDealroom has a profile for this one. Try Dealroom →, the pretraining researcher who quit Anthropic in September with a warning that the AI race is "gambling with our lives", has shared a chart he says comes from Anthropic's IPO roadshow and called it "a big factor in my decision to leave." Titled "Pre-training compute multiplier over time (code)", it plots how much more effective Anthropic's training compute has become, with completion months from July 2024 to October 2026 and the line bending upward on a logarithmic axis. "It's an insane plot and I'm surprised they made it public," he wrote. The post has been viewed more than a million times.
His reading: Because the axis is logarithmic, Coxon says the curve shows "continued superexponential algorithmic progress": efficiency gains that are accelerating, not just compounding. "Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed," he wrote, referring to Anthropic's most capable model. He added that this does not strictly imply a software-only intelligence explosion, since Anthropic has also been spending more on research, but that algorithmic improvements "are on the verge of being automated." He wants frontier development paused next year so competitors can catch up, which he says would "only harm OpenAI and Anthropic", The Decoder reported.
What the chart doesn't show: The vertical axis has no numbers, and a footnote says the placement of the data points "is based on internal approximations and is solely for illustrative purposes." Anthropic has not commented on the image, and ORIENS noted that its origin could not be independently confirmed. The footnote split readers: MIT researcher Omar Khattab argued that "solely for illustrative purposes" makes the chart literally content-free, while Thariq, who worked on Claude Code at Anthropic until October, replied that it is standard legal wording and that the graph should be taken seriously. Coxon, who resigned on 8 September after about four months at Anthropic and earlier spent time at OpenAI, said he would not break his non-disclosure agreement but wished "we could talk about the absolute numbers."
What Anthropic says on the record: Compute multipliers are a central pitch of the IPO. In the roadshow video, co-founder Tom Brown explains that a 4x multiplier means training the same model with a quarter of the compute, that multipliers "stack", and that the company believes this "can accelerate even further as we approach recursive self-improvement", with Claude eventually discovering multipliers by itself, according to a transcript published by Simon Willison. Anthropic's IPO fact sheet separately reports a more than 4x rise between Q1 2025 and Q3 2026 in what it calls compute efficiency, a revenue-per-compute measure rather than the pretraining metric in Coxon's chart. Anthropic is due to price its shares on 18 November.
Sources: Jacob Coxon on X, 13 November
The Decoder, 14 November
ORIENS, 14 November
Simon Willison, roadshow video transcript, 11 November
Anthropic IPO fact sheet
TIME, 9 September
