GPT-6 Luna

GPT-6 Luna is OpenAI's smallest, cheapest GPT-6 model — built for high-volume tasks like extraction, summarization, and routing at $0.10/M tokens.

What Changed

Released alongside GPT-6 Sol on September 22, 2026, Luna is the small, fast end of the GPT-6 family — the model OpenAI positions for extraction, summarization, request routing, and routine customer interactions rather than the hardest reasoning work. The idea isn't to run one model for everything; it's to reserve deeper reasoning (Sol, or Astra) for the steps that actually need it, and hand the high-frequency, lower-stakes steps to Luna at a fraction of the cost. At $0.10 per million input tokens, Luna is 50% cheaper than GPT-5.6 Luna on input and priced to make running the same task millions of times a non-issue.

What It's Actually Good At

• On AutomationBench (business workflows across 47 tools), Luna at high reasoning effort improves 5.4 percentage points over its predecessor GPT-5.6 Luna, while costing 58% less per task.
• On DeepSWE (real-codebase engineering tasks), Luna scores 66.6% at max effort — in line with Claude Opus 5 and Claude Fable 5 at their medium settings, but 93% and 96% cheaper per task respectively.
• On OSWorld 2.0 (computer-use tasks), Luna at max effort beats GPT-5.6 Sol at medium effort — a smaller, cheaper model outperforming the previous generation's mid-tier model — at roughly a tenth of the cost.

All of the above are OpenAI's own benchmark numbers under settings and cost assumptions OpenAI chose — useful as a directional signal, not an independently verified result.

Where It's Meant to Fit

Luna isn't trying to be the smartest model in a request — it's meant to be the one you don't think about because it's cheap enough to call constantly: tagging support tickets, pulling structured fields out of documents, summarizing a thread before it goes to a person, deciding which of a few tools to call next. OpenAI also reports substantial factuality gains at higher effort settings, saying Luna can match GPT-5.6 Sol's accuracy at roughly a hundredth of the cost on its internal factuality evaluation — again, an OpenAI-reported number based on de-identified conversations where users had flagged a prior mistake, not a representative sample of everyday accuracy.

Pricing and Access

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1.05-million-token context window and up to 128,000 tokens of output. It's the one GPT-6 Sol/Luna model with a genuinely free path: Free and Go ChatGPT users can try it in the desktop app, while Plus, Pro, Business, Enterprise, and Edu accounts get it in ChatGPT Work and Codex. It's also available via the API as gpt-6-luna. Regular ChatGPT chat doesn't have it yet.