We Reproduced Anthropic's Mythos Findings with Public Models | Hacker News

Hacker News users debated claims that Anthropic’s Mythos results were reproduced with public AI models.

Critics said the setup was not faithful: it appeared to add detailed prompts, file or line-level hints, and chunked tasks.

That may have given public models information they would not have had in Anthropic’s original evaluation.

Others questioned whether Anthropic used stronger internal resources or undisclosed prompting, making the benchmark comparison potentially unfair.