GPT-6 Sol and Luna put everyday automation in reach

OpenAI has introduced GPT-6 Sol and GPT-6 Luna, expanding its model range with options it describes as faster and more affordable than Astra. In its announcement, OpenAI reports improvements in business workflows, factual reliability and computer use compared with their predecessors. For a business with five to fifty people, the useful question is whether a routine task now deserves another look at automation.

OpenAI lists lower API prices for both models compared with GPT-5.6 promotional pricing. These are charges for the text a model processes and produces when used through an application or automation. The announcement does not establish what your complete workflow will cost. Treat it as a reason to test the economics of a specific job, rather than assume a saving across your business.

Match the model to the job you need done

OpenAI positions Astra as its strongest model overall, while Sol and Luna offer different capability and cost options. Its reported business-workflow tests suggest those options deserve attention: on AutomationBench, Sol at a high reasoning setting outperformed the tested Claude Opus 5 configuration at 9% of its cost per task. OpenAI also reports that Luna improved on its predecessor while costing less per task.

Those are evaluation results reported by OpenAI, rather than a forecast for your office. The announcement notes that its research and API evaluations may produce different results from production ChatGPT because the instructions and available tools can differ. Use the findings to choose what to trial, not to promise your team a particular result.

For your business, consider testing a cheaper model on a bounded task such as sorting incoming enquiries, extracting details from documents or drafting a weekly summary. These are trial ideas, not capabilities verified for your particular workflow by the announcement. Specify the input, the expected output and the point where someone checks the work.

Reliability deserves equal attention. OpenAI says Sol made about half as many factual mistakes as its predecessor in an internal evaluation. However, that evaluation used conversations where users had previously flagged errors, and OpenAI says they are not representative of typical use. Keep a review step in your trial rather than treating that improvement as permission to accept every answer.

Run one measured trial before expanding

Start with a task you can judge without guesswork. If you want an assistant to prepare information for Xero or MYOB, make the first trial a checked draft for review. The supplied announcement does not establish an integration with either product.

• Define a correct result. Write down what must be captured, what must be left out and when the assistant should ask for help.
• Compare the same examples. Give Sol and Luna the same representative work, then record errors, omissions and how much editing each output needs.
• Count the whole job. Include setup, model usage, checking and corrections when deciding whether to continue.
• Set an expansion rule. Agree on acceptable quality and effort before adding more tasks.

For access, OpenAI says both models are available through its API and are rolling out in ChatGPT Work and Codex for eligible paid plans, but are not yet available in Chat. Confirm the intended access route before planning your trial. Your decision should rest on a checked result from your own work, with a cost and review burden you are willing to repeat.