Gemini 4 Neon debuts at #8 on KradleBench: smart but 2.4x slower to act
nathanbenaich · x · 2026-11-21
KradleBench, a benchmark that evaluates LLM agents in simulated scenarios, reports Gemini 4 Neon landing at #8, behind Hydrogen at #7. It scores higher on raw intelligence (Creativity/Logic) but acts 2.4x slower (Conflict/Economy), leading to mishaps such as drowning at sea because its thinking loop takes too long. A demo video is available.
More from Models
- Old Gary Marcus take on LLMs resurfaces as Claude cracks hard math — inductionheads · 2026-11-21
- Researcher: the 'good data' labs buy isn't user chatbot conversations — teortaxesTex · 2026-11-21
- Extrapolating the 80% TH 1.1 chart puts Opus 5.5 at a 45-hour task horizon — ben_j_todd · 2026-11-21
- Inkling 2 Mini quantized to 5.4GB/8.9GB, keeping 91-97% performance — teortaxesTex · 2026-11-21
- Early hands-on: new model Inkling blends Gemma-like fluency with Qwen-level technical chops — Altruistic_Heat_9531 · 2026-11-21
- Ternary Bonsai-2 vs Qwen3.5-2B on A13/A14/M3: 32-55% faster decode, and why thermal logging matters — leo-k7v · 2026-11-21