OpenAI Introduced GPT-6 Sol and Luna With Lower Usage Costs

Published: September 23, 2026 – 10:33 UTC

OpenAI has released GPT-6 Sol and GPT-6 Luna — more affordable models in the GPT-6 family.
Sol’s API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens; for Luna, it is $0.10 and $0.50, respectively.
Independent tests showed a significant reduction in task execution costs, although performance gains were uneven.

OpenAI introduced the GPT-6 Sol and GPT-6 Luna models. They expand the GPT-6 family, whose flagship remains the previously released Astra.

According to the company, the new solutions were trained using the same methods applied in the development of GPT-6 Astra. OpenAI has brought some of the flagship model’s improvements in professional tasks, coding, computer use, factual accuracy, and alignment to faster and cheaper versions.

GPT-6 Sol is geared toward complex work tasks and agentic scenarios, while Luna is designed primarily for high-volume operations where speed and cost matter. The flagship Astra remains the company’s highest-performing model.

GPT-6 Sol and Luna are already available via the API. The company has also begun rolling them out in ChatGPT Work and Codex.

GPT-6 Sol and Luna Capabilities and Comparison With Other Models

One of the main changes is pricing. GPT-6 Sol costs $2 per 1 million input tokens and $10 per 1 million output tokens, versus $4 and $20, respectively, for GPT-5.6 Sol. For GPT-6 Luna, the figures are $0.10 and $0.50, versus $0.20 and $1.20 for the previous version.

OpenAI also said that GPT-6 Sol makes roughly half as many mistakes as GPT-5.6 Sol in an internal factual accuracy test. At a high reasoning level, Luna, according to the company, is comparable to GPT-5.6 Sol on this metric at about one-hundredth of the cost.

In AutomationBench, GPT-6 Sol scored 33.2% at the xhigh reasoning level. Claude Opus 5 scored 26.9% at the maximum level. By OpenAI’s estimates, Sol’s cost per task was about 11 times lower. GPT-6 Astra scored 30.3% on low.

In the DeepSWE v1.1 coding benchmark, GPT-6 Sol scored 68.8% versus 69.9% for Claude Fable 5, but, according to OpenAI’s estimates, completed the task about 80% cheaper. GPT-6 Luna scored 66.6%. The company compared this result with Claude Opus 5 and Fable 5 at a medium reasoning level.

OpenAI recorded a similar picture in the OSWorld 2.0 computer-control benchmark. GPT-6 Sol scored 60.5% versus 60.3% for Claude Opus 5 on medium, while the estimated task cost was about 80% lower.

However, these comparisons do not account for Claude Opus 5.5, which Anthropic introduced the same day.

The independent platform Artificial Analysis rated GPT-6 Sol at the top tier with 48 points on its proprietary Intelligence Index. GPT-6 Luna received 37 points. For comparison, Claude Opus 5.5 at the top tier scored 58 points and ranked first at the time of testing.

At the same time, Artificial Analysis analysts concluded that the key advantage of OpenAI’s new models is efficiency. The cost per Intelligence Index task for GPT-6 Sol fell from about $1.99 for GPT-5.6 Sol to $1.06, and for Luna — from $0.18 to $0.07.

In the Coding Agent Index, GPT-6 Sol gained two points versus its predecessor, while Luna lost two. The analysts also found improvements in some automation and programming tests, but recorded weaker results in a number of tasks tied to professional knowledge work.

What Experts and Users Are Saying on X

OpenAI CEO Sam Altman said that GPT-6 Sol and Luna significantly outperform the 5.6 generation models in intelligence, programming, computer use, and other areas. He also highlighted lower token prices and reduced costs for completing individual tasks.

In another post, Altman called the cost per completed task a more important metric than token pricing. In his view, by that measure, OpenAI’s new models currently have no competitive equivalents on the market. This is the company executive’s assessment, not the result of independent testing.

Artificial Analysis’ independent assessment was more restrained. The team posted on X that GPT-6 Sol and Luna did indeed push the price-to-performance frontier forward, but the overall Intelligence Index and Coding Agent Index scores remained close to the GPT-5.6 generation. Individual tests showed both improvements and regressions.

Product expert Peter Young, after testing, noted that GPT-6 Sol subjectively feels better than GPT-5.6. At the same time, for day-to-day work he preferred Claude Opus 5.5 and ranked this model above Fable and Astra.

AI entrepreneur Matt Shumer also said that he tested GPT-6 Sol for a while. He called the model high-quality, but for daily work he still prefers GPT-6 Astra or Claude Fable 5.1, and suggested that Opus 5.5 will be added to that list.

At the same time, developer David Kramer praised GPT-6 Luna. Among the advantages, he highlighted the combination of price, capabilities, and acceptable speed, and called the model one of the most well-balanced options on the market.

Overall, the first independent tests and feedback suggest that GPT-6 Sol and Luna primarily change the economics of using OpenAI models. A significant cost reduction comes alongside improvements in specific tasks, but there is still no noticeable across-the-board performance jump compared with the previous generation.

Сообщение OpenAI Introduced GPT-6 Sol and Luna With Lower Usage Costs появились сначала на INCRYPTED.

Source: Incrypted