OpenAI has expanded its GPT-6 model family with the release of two new API-only models: Sol and Luna. These additions join the existing Astra model, creating a three-tier structure designed to address different performance and cost requirements for professional and high-volume tasks.
The company has reduced API pricing for Sol and Luna by 50% compared to previous GPT-5.6 promotional rates. Specifically, the output price for the Luna model has been reduced from $1.20 to $0.50, representing a decrease of approximately 58%.
Performance benchmarks indicate that Sol achieves a score of 33.2% on AutomationBench 1.0.6 and 56.4% on Agents’ Last Exam, outperforming Claude Opus 5 in cost-efficiency. In coding tasks on DeepSWE v1.1, Sol scored 68.8%, while Luna reached 66.6%.
OpenAI has also introduced an improved prompt caching system. This update offers discounts of up to 90% for cached input reads, provided the shared prefixes are reused within a 30-minute window. These changes aim to reduce the volume of tokens requiring fresh processing.
The new models feature a communication style similar to Astra, prioritizing concise answers with less jargon. Internal testing shows that Sol produces approximately 50% fewer errors than its predecessor, while Luna matches previous performance levels at a fraction of the cost.
