Product leader ran a 7 model blind benchmark with opus 5

Claire Vo runs a benchmark most reviewers don’t bother with - 7 models, 6 tasks, blind scoring, no brand loyalty Opus 5 won outright Her actual verdict going in: timid, apologetic, verbose, weirdly dependent on human

YARDSep 1521

For the creator, we handle the plumbing.

Each customer gets their own copy of your agent, their own connected accounts, their own triggers. When you improve the agent, every copy updates. Data never crosses between customers. Under the hood: open-weight models that match Opus 5

Hugo MercierSep 1519

Opus 5.2 just unlocked beast mode

If this is actually opus 5.2, that’s a pretty serious jump the rocket scene looks way better than what opus 5 used to produce now i really wanna see how it compares to astra on more complex 3d

Vib3CodedSep 1516

19 system cards. Claude 2 through Mythos 5.1.

For a year and a half nobody asked the model about itself. Then they left two copies of Opus 4 alone with no task. They collapsed into what Anthropic called a spiritual bliss attractor. Nobody trained that. Then training moved the

ʞɔɐ𝘡Sep 153