OpenAI has introduced GPT-6 Sol and GPT-6 Luna as faster and cheaper alternatives to GPT-6 Astra, which remains its highest-performing model. The releases matter primarily because they reduce token and task-execution costs while carrying some of Astra’s improvements into lower-priced models.

OpenAI said it trained Sol and Luna using the same methods applied to Astra. The company incorporated some of the flagship model’s advances in professional tasks, coding, computer use, factual accuracy and alignment.

Sol targets complex work and agentic tasks, while Luna is intended primarily for high-volume operations in which speed and cost are priorities. Both models are available through OpenAI’s API, and the company has begun rolling them out in ChatGPT Work and Codex.

Pricing and benchmark results

GPT-6 Sol costs $2 per 1 million input tokens and $10 per 1 million output tokens, compared with $4 and $20, respectively, for GPT-5.6 Sol. GPT-6 Luna costs $0.10 per 1 million input tokens and $0.50 per 1 million output tokens, down from $0.20 and $1.20 for its predecessor.

OpenAI also said GPT-6 Sol made about half as many mistakes as GPT-5.6 Sol in an internal factual-accuracy test. At a high reasoning level, the company said Luna performed comparably with GPT-5.6 Sol on that measure at about one-hundredth of the cost.

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

Sol scored 68.8% in the DeepSWE v1.1 coding benchmark, against 69.9% for Claude Fable 5, but OpenAI estimated that Sol completed the task at about 80% lower cost. Luna scored 66.6%, a result the company compared with Claude Opus 5 and Fable 5 at a medium reasoning level.

In the OSWorld 2.0 computer-control benchmark, Sol scored 60.5%, compared with 60.3% for Claude Opus 5 on medium. OpenAI estimated that Sol’s task cost was about 80% lower. The comparisons did not include Claude Opus 5.5, which Anthropic introduced on the same day.

Independent assessment

Independent testing platform Artificial Analysis gave GPT-6 Sol 48 points at the top tier on its proprietary Intelligence Index, while Luna received 37. Claude Opus 5.5 scored 58 points at the top tier and ranked first at the time of testing.

Artificial Analysis found that efficiency was the new models’ principal advantage. The estimated cost per Intelligence Index task declined to $1.06 for Sol from about $1.99 for GPT-5.6 Sol, while Luna’s cost fell to $0.07 from $0.18.

Sol gained two points over its predecessor in the Coding Agent Index, while Luna lost two points. Artificial Analysis observed improvements in some automation and programming tests but weaker performance in several tasks involving professional knowledge work.

Early reactions

OpenAI CEO Sam Altman said Sol and Luna significantly surpassed the GPT-5.6 generation in intelligence, programming, computer use and other areas. He also emphasized their lower token prices and task-completion costs.

In another post, Altman argued that cost per completed task was more important than token pricing and said OpenAI’s new models had no market equivalents by that measure. That was Altman’s assessment rather than a finding from independent testing.

Artificial Analysis offered a more restrained view, saying Sol and Luna advanced the price-to-performance frontier but remained close to the GPT-5.6 generation on the overall Intelligence Index and Coding Agent Index. Its individual tests showed both gains and regressions.

Product expert Peter Young said Sol subjectively felt better than GPT-5.6, although he preferred Claude Opus 5.5 for daily work and ranked it above Fable and Astra. AI entrepreneur Matt Shumer also described Sol as high-quality but said he still preferred GPT-6 Astra or Claude Fable 5.1 for daily work and expected to add Opus 5.5 to that list.

Developer David Kramer praised Luna’s combination of price, capabilities and acceptable speed, calling it one of the market’s most balanced options.

Overall, the initial independent tests and user feedback indicate that Sol and Luna primarily change the economics of using OpenAI models. Their lower costs accompany improvements in some tasks, but testing has not shown an across-the-board performance leap over the previous generation.

Source: Incrypted