GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, half the promotional price of GPT-5.6 Sol.
GPT-6 Luna drops to $0.10 per million input tokens and $0.50 in output.
On OpenAI’s internal factuality evaluation, GPT-6 Sol makes around half as many errors as GPT-5.6 Sol and approaches Astra. At higher reasoning levels, Luna matches GPT-5.6 Sol while costing roughly one hundredth as much, according to the company.
Alignment and efficiency gains
Both models come with efficiencies to hit the token cache more often, resulting in 50% token reduction on tests run by GitHub «over the past several months,» OpenAI says.
The models are also much more aligned, with this generation of Sol and Luna posting orders of magnitude lower deception failures, broken searches and warning circumventions in OpenAI’s tests.
Performance like low effort Astra
GPT-6 Sol scored 68.8 percent on coding benchmark DeepSWE 1.1, compared with 69.9 percent for Anthropic’s Claude Fable 5 at its highest tested reasoning level. Luna scored 66.6 percent.
In other benchmarks, high effort Sol has about the performance as low effort Astra, while high effort Luna is somewhat behind low-effort Sol.
GPT-6 Sol and Luna are available from Tuesday in ChatGPT Work and Codex, and through the API as gpt-6-sol and gpt-6-luna. Luna is also ready for Free and Go users. None of the models are yet coming to Chat in the app.
GPT-5.5-instant retires on October 14, 2026, meaning there will be changes to default models in ChatGPT by then.
Read more: OpenAI’s presentation, launch post on X. TechCrunch, Reuters, and VentureBeat. Discussion on Hacker News and r/Singularity.