MazeBench: Top Models Score as Low as 1-8% on New Benchmark
patience_cave · x · 2026-11-13
A MazeBench leaderboard screenshot shows flagship models nearly flopping on this benchmark: Gemini 4 Hydrogen scores 1%, Claude Sonnet 5.5 scores 4%, and GPT-6.1 Terra scores 8%, with the top entry (gpt-6-astra) at just 16%.
The stark contrast makes it a shareable "new benchmark crushes frontier models" moment; the leaderboard's authenticity is unverified.
Related event: Top AI Models Flunk MazeBench, Best Score Just 8%(2 posts)→
More from Fun
- AI came for programmers first — and that gives everyone else time to adjust — justalexoki · 2026-11-13
- Dev builds open-source app after his $300 embroidery machine demanded a subscription — simpsoka · 2026-11-13
- Lex Fridman breaks silence: NVIDIA dropping Jensen Huang is a huge mistake for America — lexfridman · 2026-11-13
- Jensen Huang scandal spills into EA drama: utilitarianism, cheating, and power games — teortaxesTex · 2026-11-13
- Retrospective on how Sam Altman turned OpenAI's board coup into more power — zack_overflow · 2026-11-13
- As NVIDIA reels from crisis, Microsoft's outage invites brutal jokes — Kyrannio · 2026-11-13