Anthropic seems to have made a new breakthrough with their model, one so advanced that even ordinary people are afraid to use it.

https://www.anthropic.com/glasswing

Judging by benchmarks, it is noticeably stronger than Opus 4.6 in everything:

SWE-bench Verified: 93.9% vs 80.8%
Humanity's Last Exam: 56.8% vs 40.0%
GPQA Diamond: 94.6% vs 91.3%
Coding, reasoning, agent tasks — everywhere a significant lead

Essentially, this is the next generation after Opus. Anthropic keeps it closed for safety reasons — the model is too good at finding and exploiting vulnerabilities, so releasing it to everyone at once is currently risky.
They want to first test the protective filters on the new Opus, and then open access to the Mythos-class to a wider audience.

I work with Opus every day and it already handles complex tasks on its own quite often, so what comes next? 🤯