Benchmark
FrontierSui
A specialized evaluation benchmark for software engineering automation that emphasizes open-ended engineering problems. It requires models to perform multi-file modifications and architectural decision-making within real codebases, representing a higher difficulty level than DeepSWE.
Timeline (last 90 days)
Oct 4
Gemini 4 set a new record on DeepSui with 77.9% (vs. Opus 5.5's 74.2%), but scored only 55% on FrontierSui and ranked last on TerminalBench 4.0 with 57.4%.
Unverified50%