not surprised about the power user gap section. In my org we have some devs who burn their token allotment in the first day of the month, and never write code by hand, as well as older guys who refuse to touch it.
I always feel like there are gaps between benchmarks and real-world performance on all models, but especially open models as of late. I've used deepseek and kimi (albeit 2.6) extensively and while they work well for maybe 70 percent of tasks, its the last 30 they always trip themselves up on.
They all seem to be very poor at long-running tasks, maintaining context when a change spans multiple areas/layers of a codebase, and making architectural choices. That said, for making the next todo app or ai powered calorie tracker, they are just fine, and the most consumer friendly pricing.
i've been wondering recently if defense against prompt injection is more reliant on system prompt + fine-tuning and reinforcement training, or if it is simply how smart your model is.
to be honest i forgot chromebooks exist and now that you mention them i think they're the best solution for anyone whose computer is simply a medium to access a web browser