SWE-rebench adds GPT-5.5, Opus 4.7

This is the kind of benchmark update that matters more than a flashy one-off score: larger task batches make the ranking harder to game and more useful for real model comparison.

• –The March-to-May window and 110-task set should reduce variance versus smaller monthly refreshes, especially for models that swing on a handful of hard repos
• –Cursor Composer 2.5 standing out on cost-efficiency is a reminder that benchmark leadership is now about performance per dollar, not just raw pass rate
• –The contamination warnings matter: as models get newer, the exact task window and model release date become part of the interpretation, not just metadata
• –Adding GPT-5.5, Opus 4.7, and Kimi K2.6 keeps the benchmark relevant at the frontier, while the planned additions suggest SWE-rebench is becoming a rolling release cadence rather than a static leaderboard
• –For teams shipping coding agents, this is more actionable than generic coding benchmarks because it tests full issue resolution, edits, and test passing in real repositories