OpenAI released two cheaper additions to its GPT‑6 family on 22 September: Sol, a mid-tier model, and Luna, pitched at everyday tasks. Both were trained with the methods behind GPT‑6 Astra, which arrived earlier this month, and both are priced at half their GPT‑5.6 namesakes’ promotional rates — Sol at $2 per million input tokens and $10 per million output, Luna at $0.10 and $0.50.
The pricing is the story. OpenAI’s own figures put Sol at 33.2% on AutomationBench at extra-high effort, against Claude Opus 5 at maximum effort on 26.9%, at roughly a ninth of the cost per task. On DeepSWE v1.1 Sol scores 68.8%, 1.1 percentage points behind Claude Fable 5’s best result of 69.9%, at about 80% lower cost per task. On OSWorld 2.0 it reaches 60.5% against Opus 5’s 60.3%.
Treat the comparisons carefully. OpenAI ran its own models in its research environment and took competitors’ scores from published reports rather than testing them itself. It also flags that one Claude datapoint understates cost, because fallbacks to Opus 5 on around 40% of tasks are not counted.
Both models are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Luna in the desktop app for Free and Go. Neither is in Chat yet. In the API they are gpt-6-sol and gpt-6-luna.