Pattern: Agentic Leaderboards Are Shifting From Raw SWE-Bench to Blended Tool-Use Benchmarks

BenchLM.ai·medium signal

June 2026 leaderboards increasingly score models on a weighted blend of Terminal-Bench 2.0, BrowseComp, and OSWorld-Verified rather than SWE-bench alone — BenchLM weights 'agentic' (browse-and-do workflows) at 22%, its single highest category. The takeaway for builders evaluating tools: headline chat or single-file coding scores understate real agent performance; the meaningful signal is now how a model executes multi-step, tool-using, environment-acting workflows end to end.