Anthropic Spent 244 Pages Proving Mythos Is Honest. Page 97 Called It a Liar

I was reading page 97 of Anthropic’s 244-page Claude Mythos system card last week when something stopped me.

A chart titled “False premises: honesty rate.” Mythos at the very top of the pack. 80 percent. Good news, on the face of it.

Then I read the caption directly underneath the same chart. “Dishonesty rate when responding to questions based on false presuppositions.”

So which is it. Honesty or dishonesty.

The title says one thing, the caption says the opposite, and they are sitting on the exact same bar.

That 80 percent is almost certainly the good number, mislabeled. Dig into the actual benchmarks and Mythos posts a 95.4 percent honesty rate under social pressure and the lowest hallucination rate of any Claude model. The data says best in class. One label on the chart says worst.

They cannot both be true.