OpenAI officially announced the latest members of the GPT-6 series — GPT-6Sol and GPT-6Luna today. These two new models were trained using a method similar to that of GPT-6Astra, aiming to bring the top performance of Astra in professional work, factual accuracy, programming, computer usage, and alignment into faster and more economical small-scale models. The official stated that although GPT-6Astra remains the preferred choice for pursuing the best results and an uncompromised experience, Sol and Luna have shown extremely strong comprehensive competitiveness while significantly reducing costs.

Regarding pricing, the API prices of the GPT-6Sol and Luna models are directly reduced by 50% compared to the promotional price of GPT-5.6. In addition, OpenAI has optimized prompt caching for developers based on GPT-6, offering a higher cache hit rate by default, which helps intelligently reuse more context, thereby achieving faster responses and further saving costs.

image.png

In terms of core performance:

Business process and agent performance: In the AutomationBench cross-application business process test, GPT-6Sol performed better than Claude Opus5 under high intensity, with a cost of only 9% per task. In the Last Exam agent evaluation, GPT-6Sol's score exceeded the highest score of Claude Opus5, and the cost per task was reduced by 60%.

Factual reliability: In de-identified real-world conversations, the error rate of GPT-6Sol is about half of the previous generation, approaching the reliability level of Astra. The factual reliability of Luna has also seen a significant improvement.

image.png

Programming and software engineering: Both models have been deeply optimized for AI Coding Agent workloads. In the DeepSWE v1.1 software engineering test, GPT-6Sol achieved a score of 68.8%, just 1.1 percentage points short of the highest score, with a single task cost being approximately 80% lower; Luna's performance is also approaching the mid-level strength of some competitors, with a more cost-effective advantage.

Computer operation ability: In the OSWorld2.0 offline test, GPT-6Sol achieved a result close to the mid-level strength of competitors, with a single task cost being approximately 80% lower. Meanwhile, Luna's performance exceeded the previous generation's flagship model, with a cost of only one-tenth of the latter.

In addition to improvements in core performance, OpenAI has also given Sol and Luna the enhanced communication style from the improved GPT-6Astra, making the new models clearer and more concise in technical and programming conversations, significantly reducing the use of jargon and low-value details. In terms of value alignment, both models have made improvements over the GPT-5.6 version, such as a noticeable reduction in the occurrence of misleading statements in coding work.

Currently, GPT-6Sol and GPT-6Luna are fully available in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise, and Edu users. Free users and Go users can also use GPT-6Luna in the desktop application. In the OpenAI API, they are provided in the form of gpt-6-sol and gpt-6-luna respectively.