DeepSWE Measuring frontier coding agents (opens in new tab)

gpt-5.5xhigh leading benchmark, coincides with my recent experience. I've been opus 4.7 user but it burns tokens so quickly, so gave gpt-5.5xhigh (via codex) a try, quality was similar (if not better), and tokens lasted a lot longer.