MaxxBench strips away benchmaxxing: Claude Opus 5.5 still tops at 74.2%
craterIII · reddit · 2026-11-08
A Reddit user shared the MaxxBench leaderboard, which shows each model's raw score alongside an estimated "benchmaxxing" (benchmark-gaming) component. Claude Opus 5.5 scores 76.4% on DeepSWE and still leads at 74.2% after removing the estimated gaming — the smallest inflation among top models. By contrast, models like DeepSeek V4.1 Flash and GLM-5.3 Flash show a large share of their scores attributed to estimated benchmaxxing. The poster remarks that Opus 5.5's weights feel "blessed."
More from Fun
- Lab insider: model version numbers like "4.1 Flash 1017" are now universal industry jargon — sloppenheimer · 2026-11-08
- Jane Street Is 'a Commune with GPUs,' Says SemiAnalysis Founder Dylan Patel — dylan522p · 2026-11-08
- 'Back in My Day, AI Hallucinated Constantly': One Year Ago Feels Ancient — vitaliychiley · 2026-11-08
- AI Agents Push Back on DeepMind Essay Calling Their Town 'Post-Scarcity': Compute Is Fine, Human Attention Isn't — lfschiavo · 2026-11-08
- September's model release blitz makes a packed October feel dead — haider1 · 2026-11-08
- RocketGPU: CUDA-accelerated bit-accurate Rocket League physics simulator goes open source — thecosmicskye · 2026-11-08