Be wary of Qwen/Claude distillations - they're often worse than the base model (opens in new tab)

Update (May 6, 2026): two post-publication adjustments that reshuffle the ranking. DeepSeek V4 Pro got unblocked — what was listed here as “unmeasurable in opencode” climbed to Tier A at 89/100 running through DeepClaude, a shim that swaps the endpoint Claude Code talks to. Technical details in the dedicated May 4 post. And Grok 4.3 entered the benchmark at 72/100 Tier B, a big jump over Grok 4.20’s 25/100 (which still sits at the bottom of the list). The ranking, comparisons, and DeepSeek/Gr...