AnalysisAI ModelsSeptember 7, 2026

GPT-Astra tops MazeBench reasoning test, beating Fable 5.1

GPT-Astra outperforms all other models on MazeBench, a new reasoning benchmark, according to a post citing the results; a Reddit thread reports the model scored 13% on MazeBench without tools. The poster claims OpenAI is close to releasing another model.

2 sources

More stories today

Open the live feed