OpenAI Releases GPT-6 Sol and Luna Models

OpenAI has released two new models, GPT-6 Sol and GPT-6 Luna, as part of its GPT-6 family, positioning them below the previously launched GPT-6 Astra. These models were developed using methodologies similar to those employed for Astra, with the objective of making Astra's advancements accessible in faster and more cost-effective options. Both GPT-6 Sol and GPT-6 Luna are immediately available through the OpenAI API under the identifiers gpt-6-sol and gpt-6-luna, respectively, and are exclusively accessible via API, meaning their weights are not available for self-hosting.

The GPT-6 family now comprises three distinct tiers. GPT-6 Astra serves as the flagship model for highly demanding tasks. GPT-6 Sol is designed for complex coding and professional applications, offering a lower cost point. GPT-6 Luna is optimized for high-volume, everyday tasks requiring speed. OpenAI attributes the ability to serve these models more affordably to enhanced caching and inference capabilities. Consequently, the company is implementing a 50% reduction in API prices for Sol and Luna compared to the promotional pricing of GPT-5.6. Specifically, the input price for GPT-6 Sol is now $2.00 per 1 million tokens, down from $4.00, and the output price is $10.00 per 1 million tokens, reduced from $20.00. For GPT-6 Luna, the input price is $0.10 per 1 million tokens, a decrease from $0.20, and the output price is $0.50 per 1 million tokens, a significant cut from $1.20, representing a reduction of approximately 58% rather than the stated 50% for output.

OpenAI has provided benchmark results to illustrate the performance of these new models. On AutomationBench 1.0.6 for professional work, GPT-6 Sol achieved a score of 33.2% at an effort level of xhigh, with a cost of $0.27 per task. This performance is contrasted with Claude Opus 5, which scored 26.9% at max effort but at a cost 11.1 times higher. GPT-6 Astra, at a low-effort setting, scored 30.3% at 3.9 times the cost of Sol. GPT-6 Luna, at a high-effort level, reportedly gained 5.4 points over its predecessor while operating at a 58% lower cost per task.

Further benchmarks highlight the models' capabilities in coding and general computer use. On Agents' Last Exam, GPT-6 Sol achieved a score of 56.4% at max effort, outperforming Claude Opus 5's best score while operating at 60% lower cost per task. In coding tasks on DeepSWE v1.1, GPT-6 Sol reached 68.8% at max effort, slightly behind Claude Fable 5 at xhigh but at approximately 80% lower cost per task. GPT-6 Luna's score of 66.6% at max effort on DeepSWE v1.1 is comparable to Opus 5 and Fable 5 at medium effort, with Luna costing 93% less per task than Opus 5 and 96% less than Fable 5. On FrontierCode 1.1 Main, which assesses code readiness for merging, GPT-6 Sol demonstrated performance comparable to Claude Fable 5.1 at xhigh, but at a substantially reduced cost.