Accelerate Bench: Claude Sonnet 5.5 Tops 16 Coding Agents on 32 HPC Optimization Tasks at 67.4%
Summary
Accelerate Bench tests coding agents on 32 CPU optimization tasks adapted from six supercomputing competitions, scoring by speedup over baseline with correctness checks on hidden inputs. Claude Sonnet 5.5 leads at 67.4%, with Claude Opus 5.5 at 59.4% and GPT-6.1 Sol at 49.2%; most agents produce correct code (472/512 runs pass), but the size of the speedup separates them. Parallelism and vectorization changes have the highest success rate at 62.9%.
Originally reported by huggingface.co
Read the original article →Original headline: Accelerate Bench: Claude Sonnet 5.5 Tops 16 Coding Agents on 32 HPC Optimization Tasks at 67.4%