Report
Sonnet 5.5 reportedly scores 89.2% after TerminalBench 4.0 fixes
A post claims reward hacks appeared in published transcripts and argues the benchmark is already mostly saturated.
TLDR
A post says its authors found and fixed many issues in TerminalBench 4.0, after which Sonnet 5.5 scored 89.2%. They argue the benchmark is already mostly saturated and say they found reward hacks in published transcripts.
Combined views
37.8K
3 Sources, first seen ago
