FactArticle

Claude Mythos Preview worked for 60 hours total at an estimated API cost of ~$100,000, and the main human interventions were encouraging it not to give up and to find something worth publishing.

The experiment ran for 60 hours at significant cost, with humans mostly just pushing the model to persist and aim higher. ✦ AI generated

article author · Simon Willison's Weblog · 2026-07-28 · original ↗

Share card ↓Export .md Export .txt

Mythos Preview worked for 60 hours in total (~$100,000 in estimated API cost) and the main human interventions were to encourage it not to give up and 'find something that worth publishing'.

Read full article ↗excerpt · fair-use quotation

▦ Anthropic ✳ Claude Mythos

Around this claim

Mechanism · 1

LLMs tend to think it is impossible to solve cryptographic problems so they don't try; they need a good amount of prompting to persist.Anthropic researchers · Simon Willison's Weblog · conf 70%

Context · 1

The research should find something that is worth publishing — not low-hanging fruit, but genuinely hard findings.Anthropic researchers · Simon Willison's Weblog · conf 70%

This moment responds to

extends → In six weeks, Claude Mythos found vulnerabilities in Palo Alto Networks' code that would have taken five to seven years to find using traditional methods.Nikesh Arora · All-In Podcast extends → In six weeks, Claude Mythos found vulnerabilities in Palo Alto Networks' code that would have taken five to seven years to find using traditional methods, at a cost in the low millions.Nikesh Arora · All-In Podcast extends → Anthropic discovered cryptographic weaknesses using Claude with Mythos Preview, spending $100,000 on tokens with prompts that demanded 'proper research to find genuinly hard findings.'Article author (via Hacker News) · Simon Willison's Weblog