Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness

The authors report that five fresh instances of Qwen3.8-27B, working through shared files on disk, scored 92.4% against Fable 5's single-call 90.4% on 100 LiveCodeBench Hard problems. Their control, one Qwen3.8-27B call with no harness, reached 66.8% on hard LiveCodeBench. Each instance decomposes a problem and coordinates with the others through a shared filesystem holding a plan, notes, and the current solution. Critics on Hacker News say the benchmark is very simple and older, and qwen was trained after fable 5, which raises contamination concerns. One reviewer notes that the comparison spans two scaffolds, not one harness. A separate write-up estimates the harnessed 27B costs about 78% of what Fable 5 does at list prices, and says the 9% cost figure applies to Flash-Next only. For a team building on open weights, this is a hypothesis to test, not a procurement decision. Orchestration may lift a mid-size model on narrow coding tasks, but the gains depend on extra inference calls and rest on one benchmark. The result is author-reported on one benchmark with possible contamination, so test it on your own code. Measure cost per accepted task on your own workloads before changing your model stack.