The benchmark framing is the usual vendor move, comparing only against other open weights models and leading with agent numbers that don't travel outside the eval. I think the part actually worth caring about is the open training data and code which a few people already flagged.

For agents that matters more than the leaderboard slot because reliability is mostly about how well the model fits your task distribution and an open model you can finetune on your own traces beats a higher scoring closed one you can't touch. The catch is the A55. A 550B with 55B active needs real hardware, so this is an institutional base you adapt, not a local model. More useful than another sealed model that benches a point higher, just not for the reason the slide is selling.