for my work, shader code mostly, opus way way way overthought things, way overengineered, and kinda failed more often than sonnet or g2.5. for my purposes currently seems c4sonnet non max mode for the win. g2.5 very close behind. they alt on sorting where the other fails. also note that the pricing is wow-ouch w opus… at least compared to better (my exp) or equal (being overly fair to opus) answers are avail at 1/10th the price per prompt. I only did a few prompts and granted w many tool calls per (due to it’s own overthinking) and ran up a $50 tab for the ride. at most those prompts would have been $3 using c4-sonnet non max.
not complaining though, I said yes to the experiment, hoping to talk to closer-to-shader-god than c4-sonnet. instead opus seemed like an engineer I once employed. Brilliant coder and mind but would always over engineer to the point where not even god could avoid the house of cards before the thing fully worked. when you ever engineer at every stage you end up w a rocket that will neither function as a rocket non an amusement ride.
With everything code and AI – your milage may vary… greatly w an e+10 at times.