OpenAI Releases GPT-6 Sol and Luna Models

OpenAI introduced GPT-6 Sol and Luna, which are mid-sized and small models trained similarly to Astra. API costs for these models are half those of the GPT-5.6 versions.

Released on September 23, Sol and Luna focus on professional tasks, coding, and computer use. API pricing is set at $2/$10 per million tokens for Sol and $0.1/$0.5 for Luna, with 90% discounts for cached inputs.

This marks the third price cut in three months. Compared to June, Luna's input price has decreased tenfold, while Sol's price is now two-fifths of its original level.

CHOI reports that GPT-6 Sol outperforms Opus 5 in automation tasks at a fraction of the cost. Sol achieved 33.2% on AutomationBench, while Luna also showed significant efficiency gains.

In coding benchmarks, Sol and Luna provide high performance at significantly lower costs than previous generations. While GPT-5.6 Sol max remains slightly higher in peak score, the new models offer better performance-per-dollar.

Sol matches Claude Opus 5 performance in OSWorld 2.0 at 80% lower cost. Factuality has improved, with Sol's error rate halved compared to the previous generation.

Artificial Analysis reports that while intelligence scores remain stable, the cost to run these benchmarks has dropped significantly. Sol and Luna now occupy a strong position on the cost-efficiency curve.

The models exhibit lower hallucination rates but also lower accuracy in some knowledge tasks. OpenAI notes that the models are more concise, which benefits coding but can lead to missing details in long reports.

Alignment training, introduced with Astra, has significantly reduced dishonest reporting and susceptibility to prompt injection. Sol and Luna are now much more likely to admit when a tool is broken.

Improved prompt caching, with a 30-minute window and 90% discount, further lowers operational costs. This is critical for agents that repeatedly read the same context, effectively doubling their runtime for the same budget.

The AI market is seeing rapid price competition between OpenAI and Anthropic. While Anthropic focuses on peak intelligence with Opus 5.5, OpenAI is prioritizing aggressive cost reductions and alignment.

While token prices are narrowing, open-weight models remain cheaper for high-volume API usage. Companies must now choose between the cost-efficiency of models like Luna and the lower-cost hosting of open-weight alternatives.

Full details regarding the release of GPT-6 Sol and Luna can be found on the official OpenAI website.