OpenAI 发布 GPT-6 Sol 与 Luna：API 价格降 50%，单任务成本最低不足竞品一成

OpenAI Launches GPT-6 Sol and Luna: API Prices Drop 50%, Lowest Task Cost Less Than 10% of Competitors AIbase基地 Published inAI News · 6 min read · Sep 23, 2026 9 OpenAI has released the latest members of the GPT-6 series—GPT-6 Sol and GPT-6 Luna. Both models use a training method similar to GPT-6 Astra, bringing the state-of-the-art performance of Astra in professional work, factual accuracy, coding, computer usage, and alignment into faster and more economical models. OpenAI also emphasized that GPT-6 Astra remains its top model, and for an uncompromised experience, it should still be the first choice. Price Cut in Half, Cost Less Than 10% of Competitors The API price of GPT-6 Sol and Luna is reduced by 50% compared to the promotional price of GPT-5.6. More notably, the single-task cost: in the AutomationBench cross-application business process test, GPT-6 Sol outperformed Claude Opus 5 at xhigh intensity, yet its cost per task was only 9% of Opus 5's; at the same test with high intensity, GPT-6 Luna improved by 5.4% over its predecessor, and the cost per task dropped by 58%. In the Last Exam test evaluating agents, GPT-6 Sol scored 56.4% under max effort, higher than the highest score of Claude Opus 5 in this assessment, and its cost per task was reduced by 60%. Factual reliability has also improved: the error rate of GPT-6 Sol is about half that of its predecessor, approaching Astra-level reliability at a lower cost; Luna's factual reliability has also significantly improved. Performance Approaching Flagship, Longer Tasks More Efficient In programming, both models are optimized for AI Coding Agent workloads. In the FrontierCode 1.1 Main test, GPT-6 Sol showed significant improvement over GPT-5.6 Sol and achieved a level comparable to Claude Fable 5.1 xhigh at a lower cost. In the DeepSWE v1.1 software engineering test, GPT-6 Sol scored 68.8% under max effort, just 1.1 percentage points short of the highest score of 69.9% by Claude Fable 5, while the cost per task was about 80% lower; GPT-6 Luna scored 66.6% under max effort, performing close to Opus 5 and Fable 5 at medium effort, with the cost per task being about 93% and 96% lower respectively. In computer operations, Astra remains the strongest model of OpenAI, but Sol and Luna can complete related tasks at a lower cost. In the OSWorld 2.0 offline test, GPT-6 Sol achieved 60.5% under xhigh effort, close to the 60.3% of Claude Opus 5 at medium effort, with the cost per task being about 80% lower; Luna's max effort exceeded that of GPT-5.6 Sol at medium effort, with the cost being about one-tenth of the latter. OpenAI has also brought the improved communication style from Astra to Sol and Luna, stating that the new models are clearer in technical and programming conversations, use fewer terms and strange expressions, compress low-value details without losing substance. At the same time, the official has improved the prompt cache for GPT-6, providing a higher default cache hit rate, helping agents reuse more context and speed up responses; in terms of value alignment, both models have improved over GPT-5.6, with a lower incidence of misleading statements in coding work. Starting today, GPT-6 Sol and Luna are available in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can use Luna in the desktop application; they are provided as gpt-6-sol and gpt-6-luna in the API, and are not yet available in Chat. GPT-6SolGPT-6LunaGPT-6AstraOpenAI This article is from AIbase Daily Scan to view Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications. —— Created by the AIbase Daily Team © Copyright AIbase Base 2024, Click to View Source -