Ex-Anthropic Researcher: Mythic Can Retrain on Less Compute

Why this matters A researcher who lasted four months at Anthropic before quitting over a single IPO slide is not a small data point. If Coxon is right that Mythos can be retrained with a fraction of the original compute, the story Anthropic has been telling investors about its compute needs and cost curve deserves a harder look. For founders pricing their own training runs, this is a reminder that published compute figures from frontier labs are marketing documents as much as technical ones, written for an IPO roadshow, not a technical paper.

For researchers, it raises an uncomfortable question about what counts as a real moat: if retraining gets this much cheaper this fast, the advantage big labs claim from sheer scale looks shakier than the slides suggest. We'd treat Coxon's claim as a prompt to ask Anthropic for actual numbers, not take either side's framing at face value. Watch whether Anthropic responds directly to the compute figure, and whether other researchers who've left recently corroborate or contradict what Coxon is describing.