OpenAI GPT-6 Astra Scores 62% On ARC-AGI-3 Benchmark

4 Sept 2026, 1:42 am · 9d ago · 1 min read · OfficeChai

OpenAI's new model, GPT-6 Astra, has achieved a notable score of 62% on the ARC-AGI-3 benchmark, demonstrating improved reasoning capabilities on novel problem-solving tasks despite showing limited gains on standard benchmarks like the Artificial Analysis Intelligence Index compared to predecessor GPT 5.6 Sol. The ARC-AGI benchmark is widely recognized in artificial intelligence research for evaluating general fluid intelligence and abstract visual logic rather than mere memorization. This milestone highlights OpenAI's continued emphasis on advancing frontier reasoning systems, though performance across traditional conversational and general-purpose intelligence evaluations remains relatively stagnant across recent minor iterations.