Opus 5.5 scores 0.8% on CyberGym due to safety filters, DeepSeek V4.1 Flash 22.9%
A user criticizes Anthropic's safety filters: Opus 5.5 gets 0.8% on CyberGym with ~98% of tasks blocked, while open-weight DeepSeek V4.1 Flash 1070 reaches 22.9%; on CyberBench it's 55% vs 74%.
I love anthropic models and i'd rather use them for everything. that's exactly why this pisses me off that order never covered opus 5.5. which is what I tried to use, it was fable 5 and mythos 5, it got lifted in june, and opus 5.5 shipped in september I just looked this up to make sure I wasn't talking out of my ass: on artificial analysis cybergym, opus 5.5 scores 0.8% because roughly 98% of its tasks get blocked. free open-weight deepseek v4.1 flash 1070 scores 22.9%. on cyberbench, where refusals are failures, standard opus 5.5 gets 55% vs deepseek 74% your own system card calls the filters a "temporarily wider safety margin." that's your choice at this point, not the law