GPT-6 Luna

GPT-6 Luna, released September 22, 2026, was the small model of OpenAI's GPT-6 generation, launched alongside GPT-6 Sol nineteen days after GPT-6 Astra opened the line. OpenAI's API documentation called it "our most efficient model for focused, high-volume tasks", and the launch post placed the pair as the cost-efficiency end of a family Astra still led. It launched at $0.10 per million input tokens and $0.50 per million output — down from the $0.20 and $1.20 of GPT-5.6 Luna's promotional pricing, a cut OpenAI put at 50%, and a twentieth of Sol's rate in both columns — with cached input at $0.01, cache writes at $0.125, and the family's long-context rule past 272K input tokens: double on input and cache, 1.5x on output.

The launch figures were cost-per-task arguments. At max effort it scored 66.6% on DeepSWE v1.1, which OpenAI set beside Claude Opus 5 and Claude Fable 5 at medium effort at 93% and 96% less per task; at high effort it improved on GPT-5.6 Luna by 5.4 points on AutomationBench at 58% lower cost per task, and on the offline set of OSWorld 2.0 it passed GPT-5.6 Sol at medium effort at a tenth of the cost. On OpenAI's internal factuality evaluation it matched GPT-5.6 Sol at the higher effort levels at about a hundredth of the cost.

For a model sold on price it gave up little of the family's surface: the same 1,050,000-token context window as Sol and Astra, 128,000 max output tokens, text and image input, text output, and the full reasoning-effort range from none to max with medium as the default, on the Responses, Chat Completions and Batch endpoints. Its knowledge cutoff was May 18, 2026, four weeks later than Sol's. Luna was also the one GPT-6 model that reached free users at launch: Free and Go users got it in the desktop app, while Plus, Pro, Business, Enterprise and Edu users had both models in ChatGPT Work and Codex, and the API served it as gpt-6-luna.