OpenAI has introduced GPT-6 Sol and GPT-6 Luna, two new AI models that offer improved performance and substantially lower operating costs, expanding the GPT-6 family after releasing GPT-6 Astra earlier this month. The company announced 50% lower API pricing than the promotional pricing of its GPT-5.6 predecessors, along with improvements in coding, professional work, computer use, factual accuracy, and autonomous AI agents. The launch comes as OpenAI CEO Sam Altman outlined the company’s ambition to offer leading AI models across every price point and modality ahead of its September 29 DevDay event.
The new models bring several technological advances introduced with GPT-6 Astra to faster and more affordable offerings designed for developers, businesses, and individual users. OpenAI said improvements in inference efficiency and prompt caching have enabled it to reduce costs while expanding access to advanced AI capabilities.
Under the revised pricing structure, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20, respectively, for GPT-5.6 Sol. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for its predecessor.
GPT-6 Astra remains OpenAI’s flagship model for the most demanding workloads, while Sol and Luna offer alternatives for applications where computing costs, response speed, and deployment scale matter.
The company highlighted substantial improvements in professional work and autonomous AI capabilities. On AutomationBench, which evaluates AI agents performing business workflows using 47 tools across sales, marketing, operations, customer support, finance, and human resources, GPT-6 Sol achieved a 33.2% score at its highest reasoning effort setting, with an estimated cost of $0.27 per task.
In comparison, Anthropic’s Claude Opus 5 achieved 26.9% at its maximum effort setting, with a per-task cost approximately 11 times higher. Claude Fable 5.1, using Opus 5 as a fallback, achieved 31.4%, although OpenAI noted that the published cost comparison excludes the additional expense of fallback requests.
GPT-6 Sol also achieved 56.4% on Agents’ Last Exam, an evaluation of complex professional assignments spanning 55 sub-industries. According to OpenAI, the model exceeded Claude Opus 5’s highest reported score on that evaluation while operating at approximately 60% lower cost per task.
The company also reported improvements in factual reliability. On an internal evaluation based on anonymized conversations in which ChatGPT users had previously identified factual mistakes, GPT-6 Sol made approximately half as many errors as GPT-5.6 Sol. GPT-6 Luna also improved, approaching the factual reliability of GPT-5.6 Sol at higher reasoning-effort settings and at substantially lower cost.
These results reflect OpenAI’s internal testing and published competitor evaluations rather than guaranteed performance in everyday applications.
Coding represents another major area of development. On DeepSWE v1.1, which measures performance on complex software engineering assignments involving real codebases, GPT-6 Sol achieved 68.8%, approaching Claude Fable 5’s reported 69.9% at approximately 80% lower cost per task.
GPT-6 Luna achieved 66.6% on the same evaluation, demonstrating the ability to handle complex software engineering assignments at considerably lower operating costs. OpenAI reported that Luna’s cost per task was 93% lower than Claude Opus 5 and 96% lower than Claude Fable 5 in the configurations used for those comparisons.
On FrontierCode, an evaluation that considers coding correctness alongside test quality, code style, and whether submitted changes are suitable for merging into production codebases, GPT-6 Sol also demonstrated improvements over its predecessor.
Computer-use capabilities have similarly improved. On OSWorld 2.0, GPT-6 Sol achieved a 60.5% score at its highest reasoning effort setting, compared with 60.3% for Claude Opus 5 at medium effort. OpenAI reported that Sol completed these tasks at approximately 80% lower cost.
GPT-6 Luna also exceeded GPT-5.6 Sol’s medium-effort performance on the evaluation at approximately one-tenth the cost.
The company is introducing additional improvements to its prompt caching infrastructure to help developers reduce expenses associated with long conversations and autonomous agents. OpenAI now offers discounts of up to 90% on cached input-token reads, alongside new tools for monitoring cache performance and diagnosing inefficient requests.
Developers can also adjust reasoning effort and enable or disable tools without invalidating previously cached context. Explicit caching breakpoints provide additional control over which portions of a prompt are reused across requests.
According to OpenAI, GitHub has reduced the proportion of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models over the past several months, helping improve response times for GitHub Copilot.
Beyond performance and pricing, GPT-6 Sol and Luna incorporate communication improvements introduced with Astra. OpenAI said the new models produce clearer, more concise answers, use less unnecessary technical jargon, and provide more focused responses during coding and other complex working sessions.
The company also reported improvements in alignment compared with the GPT-5.6 family, including lower rates of misleading claims about completed coding work in its internal evaluations.
GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers, while Free and Go users can access GPT-6 Luna through the desktop application. Developers can access the new models through the OpenAI API using the identifiers gpt-6-sol and gpt-6-luna. The models are not yet available in regular Chat, and OpenAI is gradually expanding availability throughout the day.
The introduction of the new models also comes ahead of OpenAI DevDay 2026, scheduled for September 29 in San Francisco. The event will feature technical sessions, demonstrations, and presentations covering the company’s latest developer tools and technologies.
Following the announcement, Altman emphasized OpenAI’s intention to provide developers with models covering multiple price points and modalities, including text, coding, image generation, and video. He also suggested that the company has additional announcements planned for DevDay, describing preparations for an unusually large number of product launches.
KEY QUOTES:
“We want the OpenAI API to feature the best model at every price point and to be the best at every modality (text, code, image, video, etc). And then we want you all to come up with great ideas and build them and to get to be happy users. The best ideas will come from you all.”
“As a side note, getting ready for this DevDay is the first time I remember ever, in OpenAI history, saying ‘this is too much stuff to launch’.”
Sam Altman, CEO of OpenAI
