OpenAI has introduced GPT-6 Sol and GPT-6 Luna, two lower-cost models designed for everyday professional work, software development, and AI agents. The company says both models use methods similar to its higher-end GPT-6 Astra model, while API prices are roughly 50% lower than the promotional prices for their GPT-5.6 predecessors.
For developers, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 for input and $0.50 for output per million tokens. Input tokens are the text, data, and instructions sent to a model. Output tokens are the text or other content it generates.
The price reductions matter most for organizations running AI agents or processing long documents at scale. In these use cases, token costs can grow quickly because a model repeatedly reads instructions, context, and earlier messages.
A focus on lower-cost AI agents
OpenAI positions Sol as the stronger general-purpose option and Luna as the budget-oriented model. Astra remains the company’s top model for the hardest work, according to OpenAI, but Sol and Luna are intended to make advanced capabilities more affordable for broader use.
In its announcement, OpenAI says GPT-6 Sol makes about half as many factual mistakes as GPT-5.6 Sol on an internal test based on de-identified conversations in which users had previously flagged errors. The company also says GPT-6 Luna can match GPT-5.6 Sol at higher reasoning settings for roughly one hundredth of the cost in that evaluation.
These figures come from OpenAI’s own tests and should be read in context. The company says its factuality evaluation deliberately uses prompts likely to trigger mistakes, rather than typical user conversations. Benchmark results may also differ from performance in ChatGPT because system instructions and available tools can vary.
OpenAI reports improvements in several agent benchmarks. On AutomationBench, which tests work across business applications, the company says GPT-6 Sol at its highest reasoning setting scored 33.2% at an estimated $0.27 per task. It compares that with 26.9% for Anthropic’s Claude Opus 5 at a reported 11.1 times the cost. The benchmark covers workflows in areas including sales, marketing, finance, support, and human resources.
For software engineering, OpenAI says Sol scored 68.8% on DeepSWE v1.1, a benchmark for complex tasks in real codebases. Luna scored 66.6%. OpenAI says Sol came within 1.1 percentage points of Claude Fable 5’s reported top result while costing approximately 80% less per task. These comparisons rely partly on publicly reported competitor data.
Caching and availability
The company also updated prompt caching, a feature that lets applications reuse previously processed parts of a conversation or instruction set. OpenAI says cached input-token reads receive a 90% discount. It has added tools intended to help developers identify missed caching opportunities and preserve cached context when they change a model’s reasoning effort or available tools.
OpenAI says Sol and Luna also adopt Astra’s communication style, with shorter and clearer responses, less jargon, and fewer unnecessary details. It further reports lower rates of misleading claims in alignment tests that examine issues such as coding deception, attempts to bypass warnings, and unauthorized actions.
GPT-6 Sol and GPT-6 Luna are available through the OpenAI API as gpt-6-sol and gpt-6-luna. They are also rolling out in ChatGPT Work and Codex for paid, business, education, and enterprise users. Free and Go users can access Luna in the desktop app. OpenAI says the models are not yet available in the standard Chat interface.
Sources
- Introducing GPT-6 Sol and Luna – OpenAI
- OpenAI launches GPT-6 Sol and Luna, saying Sol makes about half as many mistakes as GPT-5.6 Sol and Luna matches GPT-5.6 Sol’s performance at ~1% of the cost – ZDNET
- GPT-6 Sol costs $2/1M input and $10/1M output tokens, and GPT-6 Luna costs $0.10/1M input and $0.50/1M output tokens, both about 50% cheaper than GPT-5.6 – The Deep View
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
