OpenAI introduces lower-cost GPT-6 Sol and Luna models
OpenAI expands the GPT-6 range
OpenAI has launched updated GPT-6 Sol and GPT-6 Luna models, extending the GPT-6 generation introduced earlier this month with GPT-6 Astra. The company says the new releases are designed to make the generation’s capabilities more efficient and accessible while cutting API prices to half the cost of the corresponding GPT-5.6 Sol and Luna models.
GPT-6 Sol is positioned for complex tasks, including coding. GPT-6 Luna is aimed at clerical, high-volume work with clearly defined objectives, such as document summarisation, information extraction and answering quick questions. OpenAI attributes the lower pricing to improvements in caching and inference.
Reliability claims and competitive positioning
OpenAI says its latest smaller models improve factual accuracy and reduce coding errors. In an internal factuality evaluation based on de-identified real-world conversations where users had flagged model mistakes, the company says GPT-6 Sol made about half as many mistakes as its predecessor. OpenAI describes that result as Astra-level reliability at a substantially lower cost.
The announcement also compares Sol and Luna with Anthropic’s leading models, including Fable and Opus, with OpenAI claiming stronger task performance. Anthropic released Opus 5.5 roughly 90 minutes before OpenAI announced the new models, underscoring the close timing of releases in the AI model market.
Availability across OpenAI products
GPT-6 Sol and Luna are available in ChatGPT Work and Codex for most paid accounts, as well as through the ChatGPT API. Luna is also set to reach the desktop app and the Free and Go user tiers. The rollout to the ChatGPT app and website is expected to occur gradually during the day.
The Luna availability plan follows expanded free access to GPT-5.6 Luna while the new generation broadens access through paid products, desktop software and the API.
What teams should assess
For organisations using OpenAI models, the stated 50% API price reduction may change the economics of workloads that require frequent calls, particularly summarisation, extraction and coding support. The company’s internal reliability claims should be assessed against each team’s own data, task definitions and quality controls before a production migration.
A practical next step is to run controlled comparisons of GPT-6 Sol or Luna against the existing model on representative workloads, measuring output quality, coding errors, latency and total API expenditure before changing production defaults.