GPT-6 Luna Review

GPT-6 Luna is the small model in OpenAI's GPT-6 family, released on September 22, 2026 beside GPT-6 Sol, and it exists for the jobs you run thousands of times a day: pulling data out of paperwork, condensing support tickets, routing and tagging. It costs 10 cents in and 50 cents out per million tokens, half the price of the GPT-5.6 Luna it replaces, with cached input at a cent, and it carries the same 1.05 million token context and 128,000 token output limit as Sol, taking text and images in and returning text. Free and Go users could try it in the ChatGPT desktop app at launch, and from October 7 it began rolling out as their model in ChatGPT itself.

OpenAI's claim is efficiency: at high effort it gains 5.4 points on its predecessor on AutomationBench while costing 58 percent less per task. A community run on OpenAI's developer forum found the same shape on small coding jobs, passing three tasks the old Luna failed or finished more slowly, at a fraction of the cost. The trouble is the competition.

Anthropic's Haiku 5.5, released two weeks later at the same price, beats it on every benchmark in Anthropic's own table: 16.4 percent on Terminal-Bench 4.0 against 39.2, 48.9 percent on an OSWorld subset against 72.4, and 29.1 percent on Chartography against 46.4. Those are a rival's numbers, but they are lopsided, and in one hands-on DevOps test Luna's Kubernetes output passed every linter and then failed at startup.

For extraction, summaries and classification at scale it is fast, cheap and easy to reach, which earns repetitive tasks and ease of use an A. For reasoning or real coding it is the wrong tool, and its rival at the same price is simply stronger, so it is a B.