The result
OpenAI introduced GPT-6 Sol and Luna on September 22. Sol is the higher-capability workhorse; Luna is the lower-cost tier. Both inherit parts of the GPT-6 training approach behind Astra, but are intended for work that must run faster and more affordably at scale.
The API list prices are $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 and $0.50 for Luna. OpenAI describes these as 50% reductions against the GPT-5.6 promotional prices. Those numbers matter most when an application sends long context or runs many agent steps every day.
What the models actually do
On OpenAI’s AutomationBench, Sol at xhigh effort scored 33.2% at $0.27 per task. On DeepSWE 1.1, Sol at max effort scored 68.8%, while Luna at max effort scored 66.6%. These are evaluation results, not guarantees for an arbitrary company workflow, but they show why the cheaper models are relevant to coding agents.
OpenAI also reports 60.5% for Sol on the offline OSWorld 2.0 computer-use evaluation at xhigh effort, close to the Claude Opus 5 medium-effort result it cites. A model that can navigate apps, write code and use tools at a lower task cost can support more repetitions and more human review within the same budget.
Where this lands
Sol and Luna are available through the API as gpt-6-sol and gpt-6-luna. OpenAI says they are rolling into Codex and ChatGPT Work, with rollout varying by plan and product. Astra remains the deeper model for the most demanding problems, so choosing the smallest model that completes the task is still the practical move.
The comparison to watch is an end-to-end task: whether an agent finishes a useful job with fewer retries, less supervision and lower total cost. Token prices alone cannot answer that. This launch makes that experiment newly worth running for teams that previously reserved advanced models for only their hardest requests.
Source published 2026-09-22. Coverage is based on the maker’s announcement and demonstration.