The new models utilize GPT-6 Astra architecture with improvements to caching and inference that allow for 90% discounts on cached input-token reads. On the DeepSWE benchmark, Sol's cost per task is roughly 80% lower than Claude Fable 5's best result, while Luna is 96% cheaper per task. OpenAI reported that Sol also outperforms Claude Opus 5 on AutomationBench at 9% of the competitor's cost per task.
