Local Qwen3.8-27B ties Claude Fable 5 on coding but costs 78%, not 9%

Open-weight Qwen3.8-27B served locally with the GVS5H scaffold scored 92.4% pass@1 on 100 hard LiveCodeBench problems versus 90.4% for Claude Fable 5. The gap is not statistically significant (p=0.88), and a run costs $47.82 against $61.11 — about 78%, not the widely repeated 9%.

• Qwen3.8-27B with manager: 92.4% pass@1 vs Fable 5's 90.4%, p=0.88
• Cost per 100-problem pass: $47.82 vs $61.11 for Fable 5
• The 9% figure belongs to Qwen3.8-Flash-Next at $5.76, not the 27B
• Single-call 27B scores 66.8%; the scaffold adds 25.6 points