GPT-6 Sol and Luna are here. Where do they fit?
Two new models have joined Astra. The useful question is which of your tasks deserves which tier—and what the price and benchmark figures actually cover.
What OpenAI announced
OpenAI positions Sol for complex coding and agent work, and Luna for focused work at high volume. Both are available in Codex and ChatGPT Work, with a gradual rollout; the announcement says they are not yet available in Chat. The API model IDs are gpt-6-sol and gpt-6-luna.
This is an official availability and product-positioning report. We have not yet run matched real tasks or checked every account’s allowance.
Read OpenAI’s announcement (opens in a new tab)Lower API cost is one kind of saving.
| Model | Input | Output | OpenAI’s positioning |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | Complex coding and agent workflows |
| GPT-6 Luna | $0.10 | $0.50 | Focused work at high volume |
API pricing is separate from a ChatGPT or Codex subscription allowance. Caching, long input, processing options and tools can change a completed task’s bill. These prices do not imply a particular number of Codex turns.
OpenAI reports improvements on its selected professional, coding and computer-use evaluations. These are vendor-reported results with specific effort settings, task sets and cost assumptions. The announcement’s competitor comparisons include Opus 5 and older Fable results; they do not establish how the models compare with the just-announced Opus 5.5 on your work.
See the tables and methodology in the original release (opens in a new tab)Run one real task three ways.
Keep the request and acceptance check fixed.
For coding, use the same repository and a test you would actually trust. For research, decide which claims must have original sources.
Review the corrections, not just the first answer.
Track revisions, human edits, time and account limits. Stop if the task has diverged; do not turn an interesting demo into a false benchmark.
This is an editorial comparison method, not a published three-model test by The Frontier Brief.