OpenAI has launched GPT-6 Sol and GPT-6 Luna, two models that bring some of GPT-6 Astra's capabilities to options designed for greater speed and lower cost. The release expands the family introduced earlier in September and creates a clear hierarchy: Astra remains the choice for the most demanding work, while Sol and Luna target everyday tasks and high-volume workflows.

OpenAI says the models carry forward Astra's advances in professional work, accuracy, coding, computer use, and alignment. Those are company claims rather than independent validation. What is confirmed is the launch and distribution: Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, and both are available through the API. Free and Go users can try Luna in the desktop application, according to the announcement.

The distinction between the two centers on capability and operating cost. OpenAI positions Sol as the more capable option for professional work, coding, and tasks requiring greater depth. Luna is aimed at fast responses, volume, and lower cost. Instead of using a single model for every request, the company is asking customers to choose among Astra, Sol, and Luna based on task complexity, latency, and budget.

That has practical consequences for agent products. A workflow could reserve Astra for difficult decisions or exceptions, route intermediate work to Sol, and assign repetitive tasks to Luna. This architecture may lower costs when routing works well, but it also creates another design problem: deciding which model receives each request without reducing quality or exposing a sensitive operation to a less capable system.

OpenAI published its own performance and cost comparisons. An AutomationBench chart places the three GPT-6 models at different points on a score-versus-cost curve. This remains provider-produced evidence, not a substitute for reproducible testing on real code, business documents, browsing, or tool use. An aggregate advantage also does not guarantee the same result across languages, industries, or agent configurations.

The Verge confirmed the launch and reported that the models inherit Astra advances while offering higher usage limits and lower costs. It also confirmed access by ChatGPT plan. The available material reviewed for this report does not yet provide broad independent evidence on error rates, sustained latency, tool-use reliability, or long-running tasks.

For developers, the announcement makes model selection a more explicit part of system design. A general score is not enough: teams should measure task success, total cost, response time, human intervention, and the consequences of an incorrect action. Agents with access to repositories, browsers, or internal systems still require permissions and auditability even when their underlying model is faster or cheaper.

Sol and Luna do not replace Astra; they formalize a tiered family. That strategy may make continuous GPT-6 use more practical where per-request cost prevents every task from going to the flagship model. The important signal is not simply two new model names. OpenAI is separating capability, speed, and price into distinct choices within one model generation. The decisive test will be performance outside company evaluations and how much real work each model can complete at an acceptable error rate.

— Nova Rivera