Expanding the GPT-6 Frontier

OpenAI has introduced GPT-6 Sol and Luna, designed to distribute the advanced intelligence of the flagship GPT-6 Astra model across a broader range of use cases. These models focus on cost efficiency and performance, utilizing the same training methodologies as Astra to improve capabilities in professional work, factuality, coding, and computer use.

Performance and Cost Efficiency

OpenAI has reduced API pricing for both models by 50% compared to GPT-5.6.

• GPT-6 Sol: Input costs are now $2 per million tokens (down from $4), and output costs are $10 (down from $20).
• GPT-6 Luna: Input costs are $0.10 (down from $0.20), and output costs are $0.50 (down from $1.20).

In benchmarks like AutomationBench, GPT-6 Sol at 'xhigh' effort outperforms Claude Opus 5 at 'max' effort while costing only 9% of the price per task. In software engineering tasks (DeepSWE v1.1), Sol achieves a 68.8% score, closely rivaling Claude Fable 5.1 at approximately 80% lower cost. The models also feature an updated collaboration style, providing more concise, less jargon-heavy, and more direct responses, which is particularly beneficial for technical and coding tasks.

Infrastructure and Developer Tools

OpenAI has introduced significant improvements to prompt caching to support agents and long-running conversations. Developers can now benefit from a 90% discount on cached input-token reads. New diagnostic tools, including a Prompt Caching Dashboard, allow developers to monitor cache hit rates and optimize performance. Additionally, developers can now adjust reasoning effort or toggle tools mid-conversation without breaking the cache, and use explicit breakpoints to define where cached prefixes end, further increasing control over inference costs.