When Google's Gemini 4 models (
@GoogleDeepMind) don't know an answer, they make one up about 4% of the time. The next best model, Muse Wonder Full (
@AIatMeta), does it 18% of the time.
Gemini 4 Carbon (4.2%) and Gemini 4 Argon (4.3%) still get about as many questions right as GPT-6.1 Terra and Kimi K3.1 Pro, which make up answers 35% and 37% of the time.
Merge Gateway lets you route each request to the model that fits it, without changing your integration.
Scores via week of Nov 9.