OpenAI GPT-6 Sol and Luna Model Overview
Chase AIgo watch the original →
the gist
OpenAI released GPT-6 Sol and Luna, two cost-optimized models designed for high-efficiency tasks, featuring a 50% reduction in input and output costs compared to the 5.6 series and improved factual accuracy.
Model Positioning and Performance
OpenAI has introduced GPT-6 Sol and Luna as cost-effective alternatives to the flagship GPT-6 Astra model. While Astra remains the top-tier model for complex reasoning, Sol and Luna are engineered to handle lower-complexity tasks with significantly higher cost efficiency. On the Frontier Code benchmark, Sol achieves a 49.3% score at $2.14 per task, while Luna reaches 42.4% at $0.11 per task. These models provide a middle ground for developers who find Anthropic's Opus and Fable models prohibitively expensive for routine operations.
Factual Accuracy and Cost Efficiency
Beyond raw performance, the GPT-6 series demonstrates marked improvements in factuality compared to the 5.6 generation. Luna’s factual error rate on low-effort tasks dropped from 36% to 27%, while Sol improved from 8.5% to 4.6%, bringing it close to Astra’s 3.9% error rate. The release also includes a 50% reduction in input and output costs across the board for the 6-series models.
Caching and API Optimization
OpenAI has updated its caching system to prevent cache resets when users adjust reasoning effort levels (e.g., switching from low to high). Previously, changing these parameters would force a full re-processing of the conversation history, incurring unnecessary costs. Developers can now utilize a new prompt caching dashboard to identify and optimize cache breakpoints, ensuring that warm cache states are maintained more reliably during long-running API interactions.