OpenAI announced GPT-6 Sol and GPT-6 Luna on September 22, 2026, pairing the models with lower listed API rates and access through selected ChatGPT plans, Codex and the API. The rates are API charges, not ChatGPT subscription prices.
What OpenAI announced and who could access the models
OpenAI positioned GPT-6 Sol for difficult professional work, including coding and agent workflows, and GPT-6 Luna as a lower-cost option for work at scale. The company said the models were trained using methods similar to those behind GPT-6 Astra.
At launch, Plus, Pro, Business, Enterprise and Edu users could access both models in ChatGPT Work and Codex. Both were also available through the API. Free and Go users could access GPT-6 Luna in the desktop app. OpenAI said neither model was available in regular Chat at launch.
API rates separate input from output
OpenAI listed these API prices in U.S. dollars per 1 million tokens. Input tokens are the text and other content sent to a model; output tokens are what it generates.
| Model | Input per 1 million tokens | Output per 1 million tokens |
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
Compared with the promotional GPT-5.6 rates listed by OpenAI, GPT-6 Sol’s input price drops from $4 to $2 and its output price from $20 to $10 per million tokens. For GPT-6 Luna, input falls from $0.20 to $0.10, a 50% reduction; output falls from $1.20 to $0.50, about 58.3%.
OpenAI also said cached input-token reads receive a 90% discount. Developers can change reasoning effort or tool availability while keeping earlier context available for cache reuse, and can set explicit cache breakpoints.
What OpenAI’s evaluations report
On DeepSWE v1.1, OpenAI reported scores of 68.8% for GPT-6 Sol and 66.6% for GPT-6 Luna, both at maximum effort.
OpenAI also reported that GPT-6 Sol made about half as many mistakes as GPT-5.6 Sol in an internal factuality evaluation. The test used de-identified conversations in which users had previously flagged factual errors; OpenAI said those conversations were not representative of typical use.
Creator-run coding tests show task-specific results
A creator-run comparison tested GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5 on three coding tasks: a one-page website, an interactive storm-tracking dashboard and a 3D browser game. The demonstrations showed side-by-side outputs and on-screen timing, token and cost metrics. In those runs, Luna was presented as the fastest and least expensive option, Sol as an intermediate choice, and Claude Opus 5.5 as producing the strongest showcased output.
A separate creator-run coding benchmark tested 24 prompts across 12 model configurations. It reported lower coding quality for GPT-6 Sol than GPT-5.6 Sol in the configurations tested, alongside faster execution and lower cost. The tasks and methods differ across these creator-run tests and OpenAI’s evaluations, so they do not yield a single like-for-like ranking.
