OpenAI says retained reasoning and compaction lifted GPT-5.6 Sol’s ARC-AGI-3 score nearly 3x, highlighting how harness design shapes AI benchmarks.