Every time I read something like this that doesn't match my experience I'm bewildered until I remember that providers play quality games and that everybody's workload and harness is different. For me GPT 6.1 Sol is tremendous value and excellent at low-level work like DSP and 3D game engines whereas Opus is more expensive and not as capable. I also don't get along with DeepSeek at all even though everybody on HN seems to think it's the second coming.

A neutral third party who runs a matrix of closed-source coding benchmarks across multiple harnesses would be helpful but that'd take a lot of time and money...