Summary
OpenAI has introduced GPT-6 Sol and GPT-6 Luna, lower-cost models designed for professional work, coding, computer use and AI agents. The models are available in the API, ChatGPT Work and Codex, with access varying by plan.
OpenAI has introduced GPT-6 Sol and GPT-6 Luna, two models positioned below GPT-6 Astra for users who need capable AI at lower cost. The company says the models bring improvements in professional work, factuality, coding, computer use and alignment, while infrastructure changes reduce the cost of serving them.
The models are available as gpt-6-sol and gpt-6-luna in the OpenAI API. They are also rolling out in ChatGPT Work and Codex, although access depends on the user’s plan.
Lower prices for API developers
OpenAI says GPT-6 Sol and Luna are 50% cheaper than the promotional pricing for their GPT-5.6 counterparts. Prices are charged per 1 million tokens, the units used to measure text and other model inputs and outputs.
| Model | Input price | Output price |
|---|---|---|
| GPT-6 Sol | $2, down from $4 | $10, down from $20 |
| GPT-6 Luna | $0.10, down from $0.20 | $0.50, down from $1.20 |
Sol is the higher-capability model of the two. OpenAI describes it as suitable for difficult work tasks and says the lower price should give users more room to iterate. Luna is intended to provide a more economical option for simpler or higher-volume workloads.
The company attributes the reduction to improvements in caching and inference. Inference is the process of running a trained model to generate an answer. OpenAI also says developers can receive discounts of 90% on cached input-token reads when applications reuse context. A new Prompt Caching Dashboard and diagnostics tool are intended to help developers measure cache use and identify missed opportunities.
Developers can change reasoning effort or enable and disable tools without breaking earlier cache reuse, according to OpenAI. This is particularly relevant to agents that work through long tasks while repeatedly referring to the same instructions, documents or application state.
Reported results in coding and computer-use tests
OpenAI reports that GPT-6 Sol improves on GPT-5.6 Sol in its internal factuality evaluation, which uses de-identified ChatGPT conversations where users had flagged factual errors. The company says Sol makes about half as many mistakes as its predecessor, while Luna also improves at higher effort levels.
In the company’s reported AutomationBench results, GPT-6 Sol at maximum reasoning effort scored 33.2% on end-to-end business workflows and cost $0.27 per task. GPT-6 Luna improved on its predecessor by 5.4 percentage points at high effort, according to OpenAI.
The models are also aimed at coding agents. On DeepSWE 1.1, OpenAI reports scores of 68.8% for GPT-6 Sol and 66.6% for GPT-6 Luna at maximum effort. The evaluation tests long-horizon software-engineering tasks in real codebases. On OSWorld 2.0 offline, GPT-6 Sol scored 60.5% at extra-high effort, compared with 60.3% for Claude Opus 5 at medium effort in the comparison reported by OpenAI.
These results are company-reported evaluations. OpenAI says its own models were tested in a research environment or through its API, which can produce different outputs from production ChatGPT because of differences in prompts and available tools. Competitor results were taken from publicly available reports, and the factuality and alignment tests were designed to create challenging conditions rather than represent typical usage.
Availability across ChatGPT, Codex and the API
GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. OpenAI says the models are not yet available in Chat and plans to roll them out gradually in ChatGPT Work and Codex.
GPT-6 Astra remains OpenAI’s highest-capability model for the most demanding tasks, while Sol and Luna extend the GPT-6 family to lower-cost use cases. The practical change for developers is therefore not only a new model release: it is the combination of lower token prices, cheaper repeated context and model choices aimed at different levels of work.