OpenAI has introduced GPT-6 Sol and Luna models, featuring significant updates to prompt caching. These improvements reduce the cost of cached input tokens by up to 90% and lower latency for developers creating agentic workflows. The new system is designed to handle long conversational contexts more efficiently than previous iterations.
The update includes higher default cache hit rates and new diagnostic tools for performance monitoring. Developers can now use explicit breakpoints to control which prompt segments are cached. Additionally, the system allows users to adjust reasoning effort levels and switch tools without invalidating the existing cache, moving beyond the previous 1,024-token prefix matching.
Pricing for GPT-6 API access is set at approximately 50% below the promotional rates of GPT-5. When combined with the 90% discount on cached tokens, these changes aim to lower the operational costs for sustained AI interactions. The caching upgrades are being rolled out across the GPT-6 API, ChatGPT Work, and Codex.
The GPT-6 series also includes the flagship variant GPT-6 Astra, which entered limited preview on September 3, 2026, and reached general availability on September 4, 2026. Earlier versions of OpenAI's caching technology previously enabled GitHub Copilot to reduce the volume of tokens requiring fresh processing by over 50% across billions of requests.
