Six-Way LLM Benchmark Compares Qwen3.8-Flash-Next and Opus 5.5 on Rust Tasks

A controlled benchmark compared Qwen3.8-Flash-Next and Claude Opus 5.5 across different harnesses and effort levels on a real Rust coding task involving LlamaStash. The evaluation ran 12 pull requests using Claude Code and Pi coding assistants, scored by multiple models and CI. The benchmark demonstrates that local models like Qwen running on consumer hardware can closely trail proprietary flagships like Claude Opus on complex software engineering tasks at a fraction of the API cost. This highlights the growing viability of local LLMs for real-world development workflows. Opus 5.5 achieved an average score of 82%, while Flash-Next scored around 66-68%, primarily trailing in edge cases, documentation, and tests rather than core functionality. Flash-Next ran locally via Halogen on an AMD Strix Halo laptop, achieving competitive speeds at essentially the cost of electricity.