Anthropic's Claude Mythos beats GPT-5.5 and prior model, AISI says
Technology
Anthropic dropped some big news: its new AI, Claude Mythos, outperformed both its previous version and OpenAI's GPT-5.5 in recent cybersecurity tests, according to a UK AI Security Institute (AISI) blog post from May 2026.
This shows how quickly advanced AI is leveling up.
Models nearly aced 2.5 million token cap
Both Mythos and GPT-5.5 nearly aced the tasks, even with a cap on how much they could process (2.5 million tokens).
The AISI blog post points out that these AIs are getting better fast (task lengths have doubled every few months since 2024), and their true potential might be even higher if testing limits were relaxed.
It's clear: evaluating just how powerful these models are is getting trickier as they evolve.