Jaeden Schafer

Anthropic has actually acknowledged that this exists.

I think what makes this really interesting is that this was in those documents.

The leak references a new model tier called Capybara.

So basically for context right now, Anthropic has Haiku, Sonnet, Opus, and those are kind of like their three from lowest to highest.

Opus is their best model.

Sonnet is kind of their medium and Haiku is their light model.

capybara is above opus so it's kind of like the the ultimate model it's even better than better than opus and it's basically an entirely new tier it's larger it's more capable and then anything that they publicly released and capybara and mythos appear to be basically the same underlying model so the benchmarks described in the document are huge compared to claude opus 4.6 which is already you know one of the strongest models available it's basically my daily driver personally

mythos is apparently scoring even higher on software coding academic reasoning and cyber security tasks the thing that i think is really important is that anthropic's own draft blog post their uh you know this is their own internal writing said that the model poses quote unprecedented cyber security risks and i think you have to sit with that for a second right this is anthropic this is the company that has built their entire identity around ai safety

They have, you know, constitutional AI.

They have, you know, responsible scaling policies.

They've positioned themselves.

It's like they go to war with the entire government basically to make their model safe and not used for like, you know, quote unquote, unsafe tasks or autonomous AI, all this kind of stuff, right?

And their own documents are flagging this model as a cybersecurity concern.

So I think that is pretty wild.

Something, you know, basically how the markets are responding to this, Bitcoin dropped, software stocks dipped.

I think this is, you know, pretty straightforward when you have a model that's dramatically better at cybersecurity.

And the capability cuts both ways.

It could be used to find and also patch vulnerabilities.

So hackers can use it.

And of