GVS5H: Five Qwen3.8-27B Models Match Claude Fable 5 on LiveCodeBench Hard (github.com)

Recent advancements in AI coding performance have been made with the introduction of ledger-based zero-shot self-orchestration. This innovative training-free approach allows multiple instances of a model to collaboratively decompose problems and manage solutions through a shared filesystem. In evaluating nine various models against the benchmarking suite LiveCodeBench, the method demonstrated significant improvements—up to 23.2 percentage points—on the latest hard problems. Notably, the orchestrated GPT-5.6-Terra achieved a pass rate of 88.0%, closely following Claude Fable 5's 90.4%, but at a mere 19% of the operational cost, while the open-weight Qwen3.8-27B model surged from 69.2% to 92.4%, slightly surpassing Fable 5. This breakthrough is particularly significant for the AI/ML community as it offers a cost-effective alternative to large proprietary models, enabling access to frontier-level coding performance without excessive financial barriers. The gains resulted from effective problem decomposition and persistent context management highlight the potential for organized inference strategies to optimize performance. While not all models benefited equally, the findings encourage further exploration into orchestration techniques that could revolutionize coding tasks within AI development, pushing forward the capabilities of self-hostable models.