OpenAI launches GPT-6 Sol and Luna: similar performance but with costs up to 60% lower

Just minutes apart, OpenAI and Anthropic unveiled new models aimed at cementing their dominance in the AI arena. On one side, OpenAI launched GPT-6 Sol and GPT-6 Luna, two models focused on efficiency and lower costs within the GPT-6 family, which already included Astra.

Luna is the ultralight variant (designed for maximum speed), while Sol is positioned as a balanced performance/cost option for high query volumes. Meanwhile, Dario Amodei’s company launched Opus 5.5, a high-end model optimized for complex reasoning, large-scale coding tasks, and agentic workflows, which has already claimed the top spot in the Artificial Analysis ranking.

Key takeaways on GPT-6 Sol and Luna

The first thing to keep in mind is that GPT-6 Sol and GPT-6 Luna are not going to become OpenAI’s new flagship models (star models, quite literally). That role will continue to belong to Astra, its most powerful model for complex tasks, which the company itself describes as “the world’s most intelligent and aligned model.”

In that sense, both Sol and Luna are aimed at occupying a lower tier, targeting users looking for efficient solutions at a lower price point. For example, in the Artificial Analysis Intelligence Index, GPT-6 Sol and Luna deliver virtually no overall score improvement over their GPT-5.6 predecessors. There are gains in some tests and setbacks in others.

What does change, dramatically, is the cost.

As OpenAI explains, “GPT-6 models lead the cost-intelligence frontier, combining exceptional capabilities at each level with infrastructure that deploys them efficiently at scale. Improvements in caching and inference allow us to offer these models at lower cost, and we are passing those savings directly on to users and customers by reducing Sol and Luna API prices by 50% compared with their GPT-5.6 promotional pricing.”

Looking again at Artificial Analysis benchmarks, we can confirm that the main generational leap with Sol and Luna lies in how much it costs to get them to do what they do. The platform describes the overall results as similar to GPT-5.6, but with a cost per task roughly 50%–60% lower.

• GPT-6 Sol (max) costs $1.06 per task, about 50% less than GPT-5.6 Sol (max) at $1.99.
• GPT-6 Luna (max) costs $0.07 per task, about 60% less than GPT-5.6 Luna (max) at $0.18.

Beyond that, these two models include some improvements shared by both, such as the communication style already used by GPT-6 Astra, something OpenAI believes will be helpful in technical and coding conversations. Sam Altman’s company promises greater clarity, less jargon, fewer odd expressions, fewer irrelevant details, and slightly shorter answers overall, without sacrificing content. Interestingly, that push toward more concise responses coincides with some setbacks detected by Artificial Analysis in complex professional tasks, where its evaluators observed deliverables that more frequently omitted required elements.

Both GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex starting today for all Plus, Pro, Business, Enterprise, and Edu users. Meanwhile, Free and Go users can also access GPT-6 Luna from the desktop app.

And now, on to Sol.

GPT-6 Sol

OpenAI makes a curious mistake (in my humble opinion) when discussing the capabilities of the new version of Sol, explaining that it delivers “greater intelligence and better results compared with competing models at a similar price point.”

And those competitors are Anthropic’s, of course.

Specifically, OpenAI says that GPT-6 Sol, at high effort, “outperforms Claude Opus 5 at max effort, at a cost per task that is just 9% of Opus 5’s.”

That would be wonderful, were it not for the fact that, as we mentioned, Claude’s new comparable model is no longer Opus 5, but Opus 5.5, and I am afraid Sol does not come off nearly as well in that comparison. If we look at the Artificial Analysis index, the leading benchmark platform of this kind, we can see that Sol falls very, very far behind the top-ranked model, which is, surprise, Opus 5.5. (Quite remarkably, Opus 5.5 even outperforms Astra, Sol’s bigger sibling.) As a result, the new Sol lands ten points behind Opus 5.5 in the Artificial Analysis Intelligence Index: 48 versus 58.

That said, OpenAI maintains that the new version of Sol makes roughly half as many errors as its predecessor, bringing it closer to Astra’s reliability at a much lower cost.

GPT-6 Luna

As mentioned, Luna will likely be the model used most by the majority of users, especially those on the Free plan. At high effort, GPT-6 Luna improves on its previous version by 5.4 percentage points, while cutting cost per task by 58%, from $0.18 to just $0.07. That said, it does not improve across the board: in the Coding Agent Index, it actually slips slightly compared with its GPT-5.6 predecessor.

Image: ChatGPT