HydraFusion cuts coding-agent cost 67%

GitHub’s HydraFusion research preview improved verified TerminalBench 2.1 quality by 4.9 percentage points while cutting estimated workflow cost 67% versus Claude Opus 5. GitHub’s benchmark report says its router chooses among single, cascade, and critique workflows, but results are offline and depend on the evaluated model pool, prices, and medium reasoning settings. Teams should test HydraFusion as a routing layer—not a model replacement—and compare cost per accepted patch against Octobench’s harness-level coding-agent evidence.