Five Qwen Models Beat Fable 5

🔥 What happened A new multi-agent system called team-solve hit 9.8/10 on LiveCodeBench Hard using five Qwen3.8-27B models – beating the original harness (9.6) and crushing Fable 5 alone (8.0). The entire codebase was generated by an AI from a method description. 💡 Why it matters The kicker: a simple shell script decides success, not a model. 49 of 50 runs solved, total cost: $35.79 for 15.3M output tokens. This proves multi-agent pipelines are dirt cheap today – and that orchestration matters more than the base model. ⚡ Our take The real story isn't the score – it's that nobody hand-wrote the files. If a 27B model can build a better harness than humans, we should start questioning our own jobs.