OpenAI has expanded its GPT-6 line-up with two new models, Sol and Luna, joining GPT-6 Astra. The pair splits the workload by use case rather than by capability tier alone: Sol is built for complex coding, heavy reasoning and agentic workflows, while Luna is tuned for high-volume, low-latency, low-cost tasks.
Both inherit training improvements first introduced with Astra, including gains in coding accuracy, factual reliability and computer-use tasks. The headline change, though, is price. Sol costs $2 per million input tokens and $10 per million output tokens through the API; Luna costs $0.10 and $0.50. Several reports put this at roughly half the cost of the equivalent GPT-5.6 models, which makes GPT-6-class reasoning considerably cheaper to run at scale than it was a generation ago.
Access runs through ChatGPT Work, Codex and the API, with enterprise administrators needing to switch the models on before their teams can use them; neither is yet available in the standard ChatGPT consumer tier. Luna is the exception on the consumer side, reaching Free and Go users inside the desktop app, which fits its role as the cheap, fast option OpenAI wants deployed widely.
The model identifiers are gpt-6-sol and gpt-6-luna. By releasing two specialised models rather than one general one, OpenAI is following the same split its rivals have adopted: a capable but costly model for hard problems, and a cheap, fast sibling for everything that does not need it.
Source: OpenAI
