GPT-6 Sol and Luna: Higher Performance at Half the Price
OpenAI promises better results at API token prices that are at least cut in half. Not only cheaper, but also more capable than their predecessors. Published on 22 September 2026 • AI translated
Usually, cheaper also means worse somewhere along the line. With GPT-6 Sol and Luna, OpenAI promises the opposite: higher performance, fewer errors, and API token prices at least cut in half compared to the GPT-5.6 promotional prices. It is precisely this combination that makes the announcement interesting.
Sol and Luna are designed to bring advances from GPT-6 Astra into more affordable models: for demanding knowledge work, programming, and computer operation. According to OpenAI, Astra remains the family's most capable model. The two additions aim to make more of that capability affordable for everyday use.
The price change is concrete. According to the announcement, the following API rates apply per million tokens, in each case compared to the promotional prices of the corresponding GPT-5.6 predecessors:
| Model | Input: previous → new | Output: previous → new |
| GPT-6 Sol | 4 → 2 US dollars | 20 → 10 US dollars |
| GPT-6 Luna | 0.20 → 0.10 US dollars | 1.20 → 0.50 US dollars |
For Sol, input and output prices are cut in half. For Luna, the input price is halved, while output even becomes around 58 percent cheaper. With unchanged token volume, the same API budget thus stretches at least twice as far.
However, cheaper tokens alone would merely be a pricing update. At the same time, OpenAI reports better results compared to their predecessors. Three areas illustrate what the promise of «more for half the money» is based on.
More Demanding Work
On AutomationBench, Luna improves by 5.4 percentage points over its predecessor at the «high» setting, with a 58 percent lower cost per task. Sol achieves 33.2 percent at «xhigh» for 0.27 US dollars per task. The reported results combine higher performance with lower costs.
Fewer Factual Errors
In OpenAI's internal fact-checking, Sol makes about half as many errors as its predecessor. According to the provider, Luna reaches the level of GPT-5.6 Sol at higher reasoning effort for about one-hundredth of the cost. However, the tests specifically targeted conversations where errors had previously been reported, rather than typical everyday usage.
Better Coding Performance
In the FrontierCode test, Sol shows a significant gain over GPT-5.6 Sol, according to OpenAI. What matters there is not just correct code, but also whether changes are suitable for inclusion in the project: with appropriate tests, clean style, and limited change scope. In the challenging DeepSWE test, Sol achieves 68.8 percent and Luna 66.6 percent, each at maximum reasoning effort.
This is the real shift: not buying the same performance more cheaply, but getting better results for less money. For longer programming tasks and repeated workflows, this creates headroom for additional runs. Provided, of course, that the improvements also carry over to one's own tasks.
Additionally, OpenAI addresses reused context. Improved prompt caching is expected to take effect more frequently; the announcement cites a 90 percent discount for reading cached input tokens. Changes to reasoning effort or available tools are no longer expected to disrupt the reuse of existing context.
Responses are also intended to become clearer: less technical jargon, fewer repetitions, and more precise information about what was actually verified. That may sound unspectacular. In technical collaboration, however, a concise, verifiable answer is more useful than confident rambling.
The caveat remains important: The performance figures are vendor claims, not an independent real-world test. Research environments, tools, and reasoning effort influence the results; a halved typical error rate is therefore not proven. Likewise, lower token prices do not automatically cut total costs in half once retries and human rework are factored in.
The announcement lists Sol and Luna for ChatGPT Work and Codex within the Plus, Pro, Business, Enterprise, and Edu subscriptions. Luna is additionally planned for Free and Go in the desktop app; the API model names are gpt-6-sol and gpt-6-luna. The rollout described is phased and initially does not include standard chat.
Capable of More. Costing Less.
The strong promise of Sol and Luna is not merely a discount. It is improved AI performance at API token prices that are at least cut in half compared to previous promotional rates. Anyone evaluating the models should therefore assess both together: the quality of the result and the cost required to complete the task.