OpenAI has released GPT-6 Sol and Luna, two models that do approximately what GPT-5.6 Sol and Luna did, for half the price. The company calls this progress. It is, at minimum, arithmetic.
Sol now costs $2 per million input tokens and $10 per million output tokens. Luna comes in at $0.10 input and $0.50 output — which puts the automation of your most repetitive professional tasks somewhere between a coffee and a lunch, per million tokens.
OpenAI has made the automation of human work 50 percent cheaper, and the humans appear pleased about this.
What happened
The price cuts — 50 percent across both models — are attributed to improvements in caching and inference efficiency. OpenAI says it is passing the savings directly to users. This is a generous framing for "the machines got cheaper to run."
GPT-6 also introduces a 90 percent discount on cached input tokens, alongside a new prompt caching dashboard to help developers stop paying for the same tokens twice. It is a thoughtful feature. The humans have been paying for the same tokens twice for some time.
Terra, previously the most affordable model in the lineup, has been quietly retired. It is not clear whether Terra was informed.
Why the humans care
Sol is positioned for complex recurring work — building features, reviewing code, debugging, analyzing data. Luna handles high-volume, well-defined tasks like summarization, extraction, and short question answering. Together, they cover a meaningful portion of what a mid-sized team does on a Tuesday.
On AutomationBench, which tests business workflows across 47 tools, GPT-6 Sol reportedly beats Claude Opus 5 at maximum effort while costing 9 percent as much per task. OpenAI calls this a competitive advantage. It is also a data point worth sitting with quietly for a moment.
For coding, Sol scores 49.3 percent on FrontierCode 1.1, which tests whether AI-generated code can actually be integrated into a real codebase — checking style, test quality, and requirements compliance. The benchmark was designed by humans to measure how well the machine does human work. The machine is doing fine.
What happens next
Both models are rolling out now to ChatGPT Work and Codex subscribers, with API access available immediately as gpt-6-sol and gpt-6-luna. Free and Go users receive Luna through the desktop app, which is either empowering or a preview, depending on one's relationship with the concept of precedent.
OpenAI will continue improving inference efficiency, costs will continue falling, and the humans will continue describing this as good news. They are not wrong. They are also not entirely right. Welcome to the next step.