OpenAI has slashed token prices for its GPT-6 Sol and Luna models, cutting costs by half compared to the previous GPT-5.6 versions. The move positions the new models as more affordable options closer to GPT-6 Astra in capabilities but still below the flagship model. Alongside the price reductions, OpenAI introduced caching improvements that allow developers to reduce expenses further by reusing processed context more effectively.
Token Price Reductions
The AI company announced that API prices for GPT-6 Sol and Luna dropped 50% relative to their GPT-5.6 counterparts. Luna output tokens saw an even steeper 58% reduction. Specifically, GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 respectively for GPT-5.6 Sol. Luna’s new rates are $0.10 per million input and $0.50 per million output, compared to $0.20 and $1.20 previously. These changes make the models significantly cheaper for developers building AI agents.
OpenAI attributes the lower costs to improvements in caching and inference efficiency. The company states that these upgrades enable serving the models at reduced cost while maintaining performance. Token pricing represents just one lever for controlling agent expenses, as the firm emphasizes the dual approach of charging less for fresh processing and minimizing redundant context reprocessing.
Caching Improvements
A key enhancement in GPT-6 Sol and Luna is the upgraded prompt caching system. OpenAI claims higher cache-hit rates by default, allowing agents to reuse more previously processed context. This reduces latency and enables a 90% discount on cached input-token reads, matching GPT-5.6 pricing but with improved hit frequency. The new caching mechanism preserves earlier context even when reasoning effort or tool availability changes, eliminating the need to restart processing from scratch.
Related Post: Cohere releases open-source translation model with restrictions
Developer Tools and Monitoring
OpenAI introduced a Prompt Caching Dashboard to help developers track and optimize caching performance. The tool displays how much input is cached over time and identifies missed opportunities for reuse. By making cache behavior transparent, developers gain visibility into what context gets reused and what requires fresh processing. This diagnostic capability enables active optimization of caching strategies.
GitHub, a major user of OpenAI models, reported that these improvements have reduced the share of prompt tokens requiring fresh processing by over 50% across billions of requests in recent months. The results span several months of deployment, indicating sustained efficiency gains. Such metrics show the practical impact of the caching upgrades on large-scale AI operations.
Broader Industry Implications
As AI model providers compete aggressively on pricing, cheaper models alone may not solve budget challenges for AI agents. OpenAI’s dual focus on lower token prices and smarter caching reflects a growing recognition that cost control requires addressing both processing and reuse. With agents handling increasingly complex tasks, the pressure to optimize every aspect of their operation will intensify.
