The fact that claude and gpt 5.5 have nearly the same scores tells me your benchmark is not capturing a significant gap in capability between these two. What the linked page says about Claude is true in my experience: It frequently forgets important instructions and likes to take lazy shortcuts. Gpt by contrast is much more attentive and takes its time when needed to deliver a complete and robust solution. I have tested both models on two private repos (c#, go) on two long-horizon tasks with well-defined stop conditions and observed the same pattern in both cases. Both models still require a large harness to reduce shortcuts and architecturally unclean code, but gpt performs much better, to the point where I find claude unusable for any significant work.