Tag
ProgramBench results show Fable 5 achieving double the performance of Opus 4.8, even with fallback to 4.8 in 99% of runs.
A study from the Opus 4.8 card shows that while multi-agent systems do not achieve better results on ProgramBench, they reach mediocre solutions twice as fast.
GPT5.5 achieved the first solve on the difficult ProgramBench SWE benchmark, significantly outperforming Opus 4.7.