The race among frontier AI labs is no longer only about who builds the smartest model. Increasingly, it is about who can deliver strong performance at a price businesses can afford to run millions of times a day. OpenAI’s latest release makes that shift impossible to ignore. With GPT-6 Sol and GPT-6 Luna, the company has brought much of the capability of its flagship model into two lighter, cheaper options — and cut API prices roughly in half in the process.
In this article, we break down what the two models are designed for, how much they cost, where you can use them, and why independent observers are urging developers to read the benchmark charts carefully before switching.
What Are GPT-6 Sol and GPT-6 Luna?
OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22, 2026, a few weeks after it launched GPT-6 Astra, the most capable model in the new generation. Rather than being stripped-down versions of Astra, the company says both models were trained using similar methods, with the goal of passing Astra’s strengths down to models that are quicker and less expensive to operate.
The two models occupy different tiers:
- GPT-6 Sol is the workhorse for demanding jobs — complex coding, autonomous AI agents, and business workflow automation.
- GPT-6 Luna is built for high-volume tasks with a clear goal, such as summarizing documents, pulling information out of files, or answering short questions quickly.
Astra remains OpenAI’s top choice for computer-use tasks, where a model operates software on a user’s behalf. Sol and Luna, by contrast, are pitched on efficiency: doing most everyday work well while consuming far fewer resources.
The Headline Feature: Prices Cut by 50%
The most immediate change for developers is cost. OpenAI priced the new models at about half of what the previous GPT-5.6 equivalents charged, per million tokens:
- GPT-6 Sol: $2 input / $10 output (down from $4 / $20 for GPT-5.6 Sol)
- GPT-6 Luna: $0.10 input / $0.50 output (down from $0.20 / $1.20 for GPT-5.6 Luna)
OpenAI credits the reduction to improvements in caching and inference efficiency, and reportedly told VentureBeat that these are permanent prices rather than a limited launch promotion. For companies running AI at scale — customer support bots, document pipelines, coding assistants — a 50% cut can be the difference between an experiment and a product that makes financial sense.
Independent testing from Artificial Analysis supports the efficiency story. On its Intelligence Index, running Sol at its maximum setting cost around $1.06 per task, compared with roughly $1.99 for GPT-5.6 Sol. Luna’s cost dropped from about $0.18 to $0.07 per task.
Fewer Mistakes, Better Code
Beyond price, OpenAI is emphasizing reliability. According to the company’s internal factuality tests, GPT-6 Sol makes roughly half as many factual errors as its predecessor, bringing it close to Astra-level reliability at a much lower cost.
Coding is another focus area. On FrontierCode 1.1, a benchmark that checks whether AI-written code is good enough to merge into real codebases, OpenAI reports that Sol improved significantly over GPT-5.6 Sol and performed on par with Anthropic’s Claude Fable 5.1, while costing less. The company also claims that on AutomationBench — a test of workflows that span multiple apps — Sol beat Claude Opus 5 at a fraction of the cost per task.
Luna arguably saw the bigger jump relative to its tier. OpenAI says the top configuration of Luna can outperform the medium setting of GPT-5.6 Sol at around one-tenth of the cost, which could make it a compelling default for routine workloads.
Read the Benchmarks With Care
As with any launch, the numbers come mostly from OpenAI’s own evaluations, and analysts have pointed out a few details worth noting:
- Some launch comparisons were made against Claude Fable 5 rather than Claude Opus 5, even though OpenAI’s own chart showed Opus 5 scoring higher on one coding test.
- Higher “effort” settings did not always produce better results. On AutomationBench, Sol at maximum effort scored slightly lower than Sol at the “xhigh” setting while costing more — a reminder that turning every dial up is not automatically the best choice.
- “Pro” variants of both models appeared on OpenRouter on launch day without being described in the announcement, so the differences between them and the base models are not yet documented.
None of this means the improvements are not real. It simply means teams should run their own tests on their own tasks before making a full migration.
Timing: A Direct Answer to Anthropic
The launch timing raised eyebrows across the industry. Anthropic released Claude Opus 5.5 on the morning of September 22, and OpenAI announced Sol and Luna roughly 90 minutes later. OpenAI repeatedly positioned the new models as outperforming Anthropic’s Fable and Opus lines on many tasks — a clear sign that the two labs are competing head-to-head for the same enterprise and developer customers.
For buyers, this rivalry is good news. Rapid releases and aggressive pricing put pressure on every provider to deliver more capability per dollar.
Where Can You Use GPT-6 Sol and Luna?
At launch, both models became available through:
- ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu plans
- The OpenAI API, using the model names
gpt-6-solandgpt-6-luna - The ChatGPT desktop app, where Free and Go users can access Luna
OpenAI noted that the models were not immediately available in standard ChatGPT chat, and enterprise administrators may need to enable them before teams can use them in a workspace. In early October, OpenAI also published an update covering safety details for the Sol and Luna models.
Who Should Switch?
If you already run GPT-5.6 Sol in production, testing GPT-6 Sol is an easy decision: the price is lower and the reported scores are higher. Developers building agents or coding tools are the most obvious beneficiaries.
Luna deserves attention from anyone processing large volumes of text — support tickets, invoices, contracts, product reviews, or research notes. At ten cents per million input tokens, tasks that were once too expensive to automate may now be worth revisiting.
Frequently Asked Questions
Is GPT-6 Sol better than GPT-6 Astra?
Not overall. Astra is still OpenAI’s most capable GPT-6 model, especially for computer-use tasks. Sol is designed to deliver much of that capability at a lower price and higher speed.
Are the new prices temporary?
According to reporting from VentureBeat, OpenAI described them as permanent rather than a launch promotion.
Can free users access these models?
Free and Go users can use GPT-6 Luna through the ChatGPT desktop app. Sol requires a paid plan or API access.
The Bottom Line
GPT-6 Sol and Luna show where the AI market is heading: strong capability is becoming cheaper, faster, and more widely available. Halving API prices while improving accuracy is a meaningful step for businesses that want to scale AI without scaling costs at the same rate. The smart move is to treat the launch claims as a starting point, benchmark the models against your own workloads, and let real results decide.