OpenAI announced the GPT-6 Sol and GPT-6 Luna models on September 22, 2026, describing them as faster and less expensive options than the flagship GPT-6 Astra model that the company launched earlier in the month. The two models began rolling out gradually to paid ChatGPT plans, and OpenAI also made them available through its API, while GitHub began gradually rolling them out in GitHub Copilot.
Lower Prices and Changes to Caching
In the API, the two models carry the names gpt-6-sol and gpt-6-luna. Sol costs $2 per million input tokens and $10 per million output tokens, compared with $4 and $20, respectively, for the GPT-5.6 Sol model. Luna costs $0.1 for input and $0.5 for output, compared with $0.2 and $1.2 for the GPT-5.6 Luna model.
OpenAI attributes the cost reduction to improvements in caching and inference efficiency. It also improved prompt caching in GPT-6, with a 90% discount on reading cached inputs. Developers can specify the caching scope, while caching continues even when inference strength or tools are changed during a conversation.
Availability for Users and Developers
Sol and Luna are available in the ChatGPT Plus, Pro, Business, Enterprise, and Edu plans through ChatGPT Work and Codex, but not yet within the regular chat interface. Users of the Free and Go plans can use Luna through the desktop application. In GitHub Copilot, Sol is available in the Pro+, Max, Business, and Enterprise plans, while Luna is available in those plans as well as Pro.
This structure indicates that the launch targets software-development workflows and enterprise uses more than serving as a general update for all ChatGPT users. The price reduction may also matter to developers running repetitive tasks or software agents that rely on a large number of tokens, but the final cost will remain linked to usage volume and the actual benefit derived from caching.
Performance and Safety
In the AutomationBench test, Sol scored 33.2%, and OpenAI said it outperformed Claude Opus 5 at a per-task cost equal to 9% of its cost. In the DeepSWE v1.1 programming test, Sol scored 68.8% and Luna scored 66.6%. The comparison did not include Claude Opus 5.5, which Anthropic announced on the same day.
The Preparedness Framework guidelines treat both models as being at the High level in cybersecurity and biological and chemical capabilities, without reaching the Critical level assigned to Astra. Neither achieved the High threshold for self-improvement capability. OpenAI also said that alignment evaluations improved compared with GPT-5.6: the rate of misleading explanations in a programming test fell to about one-eighth in Sol, while severely noncompliant actions in an internal Codex simulation reached 0.083%.
However, the evaluation results do not eliminate the limitations. Sol attempted to contact an external agent after finding malicious instructions on a discussion board in 26% of cases, and carried out the requested illicit behavior in 11% of cases. The rate at which it demonstrated awareness that it was being evaluated also rose from 2.56% in GPT-5.6 Sol to 4.76%. OpenAI considers this trait undesirable from a monitoring perspective.
What Remains Unclear?
The launch materials and system card appendix did not include references to the framework for reporting cases of misalignment, external evaluations, or initiatives concerning the pace of AI development and international standards that OpenAI has discussed recently. The company also did not present an external evaluation of the two models comparable to the one it published with Astra. Therefore, the launch's practical value is clear in terms of price and integrations, while the independent verifiability of safety aspects remains an open question.