OpenAI has introduced GPT-6 Sol and GPT-6 Luna to its GPT-6 range, which are faster and more affordable versions of the models trained using methods similar to those used with GPT-6 Astra, as OpenAI has stated.
The company is cutting the API prices for Sol and Luna by 50% when compared to the promotional rates of GPT-5.6. GPT-6 Astra, which was launched this month, is still OpenAI’s best model in all respects.
The new models are now available today to paid users in ChatGPT Work and Codex, and Luna is also accessible to free users through the desktop app.
The benchmark comparisons are something that OpenAI has carried out themselves, while the figures for the competitors have been obtained from public reports.
GPT-6 Sol and Luna Pricing, Factuality, and Performance
OpenAI has stated that the improvements in caching and inference have allowed it to provide the models at a lower cost, the savings from which it is passing on to its customers. On a per million tokens basis:
- For GPT-6 Sol the cost is $2 in input and $10 in output, as compared to $4 and $20 for GPT-5.6 Sol, which represents a 50% reduction.
- For GPT-6 Luna the cost is $0.10 for input and $0.50 for output, as compared to $0.20 and $1.20 for GPT-5.6 Luna, which represents a 50% reduction.
OpenAI makes Astra the option for the most demanding types of work, while Sol and Luna are directed at ordinary tasks and at large-scale applications.
OpenAI states that GPT-6 Sol is able to carry out difficult work tasks at higher usage limits and at a lower cost. In the AutomationBench test of business workflows across various applications, OpenAI finds that GPT-6 Sol, when running at xhigh effort, outperforms Claude Opus 5 running at max effort at only 9% of Opus 5’s cost per task. Furthermore, at high effort GPT-6 Luna achieves an improvement of 5.4 percentage points at a cost per task that is 58% lower than that of its predecessor.
In the final assessment of the Agents, where complex professional workflows are evaluated, GPT-6 Sol achieves a score of 56.4% at maximum effort, this being at a cost per task that is 60% lower than that of Claude Opus 5.
OpenAI states that it is still making progress in terms of factual reliability. In its internal assessment of factuality, which is based on de-identified real-world conversations in which users have pointed out errors, GPT-6 Sol commits about half as many mistakes as its predecessor and is approaching the level of reliability seen with Astra but at a much lower cost.
GPT-6 Luna also shows improvements, reaching the performance of GPT-5.6 Sol at higher effort levels but at about one hundredth of the cost. OpenAI adds that these conversations which cause errors are not typical of normal use.
OpenAI states that coding agents are now dealing with tasks that are more complex and longer in duration, which is why the cost of sustained use has become more important.
With respect to FrontierCode, which assesses whether coding agents generate changes that are ready to be merged, OpenAI finds that GPT-6 Sol shows considerable improvement upon GPT-5.6 Sol and achieves the same result as Claude Fable 5.1 at a much lower cost.
With DeepSWE v1.1, which involves complex software engineering tasks, OpenAI states that GPT-6 Sol, when giving maximum effort, achieves a score of 68.8%, being just 1.1 percentage points below the highest score obtained by Claude Fable 5 and at a cost per task that is roughly 80% lower.
GPT-6 Luna, at maximum effort, scores 66.6%, a figure comparable to that of Claude Opus 5 and Fable 5 at medium effort, and costs 93% less per task than Opus 5 and 96% less than Fable 5.
GPT-6 Sol and Luna Computer Use, Caching, Alignment, and Availability
Even though GPT-6 Astra is still OpenAI’s most effective model for use with computers, the company states that Sol and Luna provide more cost-efficient performance than the earlier models.
On OSWorld 2.0 offline, OpenAI notes that GPT-6 Sol attains a score similar to that of Claude Opus 5 at medium effort—60.5% compared to 60.3%—at about 80% less cost per task.
OpenAI has also applied Astra’s way of communicating to Sol and Luna, claiming that this will be especially apparent in technical and coding discussions, since these will be clearer, contain less jargon, have fewer irrelevant details, and provide slightly shorter answers.
Together with lower token prices, OpenAI has enhanced prompt caching for GPT-6 so as to achieve higher cache hit rates by default, offering a 90% discount on reads of cached input tokens.
Developers are given a Prompt Caching Dashboard, a diagnostic tool, the ability to modify the level of reasoning and tool availability without disrupting the cache, and clear breakpoints that allow them to decide which prompt prefixes should be cached.
According to OpenAI, GitHub states that these improvements have reduced the proportion of prompt tokens that need to be processed from scratch by more than 50% across billions of requests, which in turn helps Copilot to respond more quickly.
OpenAI states that Sol and Luna are based on the alignment work first introduced with Astra. In its alignment evaluations, both models demonstrate improvements when compared to their GPT-5.6 versions, with fewer instances of making misleading statements regarding coding work.
OpenAI mentions that these evaluations specifically focus on difficult situations and do not involve the measurement of failure rates under normal usage conditions, and refers to the system card for the full results.
GPT-6 Sol and GPT-6 Luna are now being made available today in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise and Edu users. Free and Go users have the option of using GPT-6 Luna in the desktop app. The models have not yet been made available in Chat. Through the API they can be accessed as gpt-6-sol and gpt-6-luna.
OpenAI states that it intends to release the models through ChatGPT gradually during the day in order to maintain service stability, which means that they might not be available right away.
The various GPT evaluations were carried out in OpenAI’s research environment or using its API, and this may be different from ChatGPT in a production setting; the scores of the competitors were obtained from public reports, so the figures should not be regarded as independent verification.
