# GPT-6 Luna GPT-6 Luna is the smallest and cheapest model of [[OpenAI]]'s GPT-6 family, below [[GPT-6 Sol]] and [[GPT-6 Astra]]. It was released on 22 September 2026 together with Sol. OpenAI describes it as its "most efficient model for focused, high-volume tasks". It replaces [[GPT-5.6]] Luna at half the price: $0.10 per million input tokens and $0.50 per million output tokens. [[Simon Willison]] noted that only two far weaker models ever came cheaper from OpenAI: GPT-4.1 Nano ($0.10 / $0.40) and GPT-5 Nano ($0.05 / $0.40). What matters is the ratio. Luna costs one twentieth of Sol per token and one hundredth of Astra, while scoring close to models that cost ten times more on several agentic benchmarks. At that price, the useful question stops being "is it the best model?" and becomes "is it good enough for this step?". ## Specifications and pricing | | GPT-6 Luna | |---|---| | Model ID | `gpt-6-luna` | | Input / output | $0.10 / $0.50 per million tokens (GPT-5.6 Luna: $0.20 / $1.20) | | Cached input / cache write | $0.01 / $0.125 | | Long prompts (> 272K input tokens) | 2× input and cache rates, 1.5× output, for the whole request | | Batch and Flex | 50% of standard rates. Fast mode: 2× | | [[Context Window\|Context window]] | 1,050,000 tokens (922,000 max input) | | Max output | 128,000 tokens | | Knowledge cutoff | 18 May 2026 (a month later than Sol) | | Modalities | Text and image in, text out | | Reasoning effort | `none`, `low`, `medium` (default), `high`, `xhigh`, `max` | Luna shares Sol's API surface: same built-in tools through the Responses API, and function calling in Chat Completions only with reasoning effort `none`. In ChatGPT, Luna shipped in ChatGPT Work and [[OpenAI Codex|Codex]] for paid plans. Free and Go users can use it in the desktop app. ## What OpenAI claims From OpenAI's launch post: - **AutomationBench**: 5.4 points better than GPT-5.6 Luna at high effort, at 58% lower cost per task - **DeepSWE v1.1**: 66.6% at max, comparable to [[Claude Opus 5]] and [[Claude Fable 5]] at medium effort, at 93% and 96% lower cost per task respectively - **OSWorld 2.0 offline**: Luna at max beats GPT-5.6 Sol at medium for a tenth of the cost - **Factuality**: at higher effort levels, it matches GPT-5.6 Sol at about a hundredth of the cost ## What independent testing found [[Artificial Analysis]] says Luna halves cost but does not raise the bar: - **$0.07 per task** at max on its Intelligence Index, about 60% less than GPT-5.6 Luna ($0.18), even though Luna uses more output tokens per task (51k vs 41k) - **Coding Agent Index: 41, down 2 points** from GPT-5.6 Luna. SWE-Atlas-QnA drops from 49% to 44%, DeepSWE v1.1 from 66% to 64% - **Hallucination rate** on AA-Omniscience drops from 93% to 77%, with accuracy roughly unchanged (44% vs 43%) - **Regressions in knowledge work**: about 75 Elo points lost on GDPval-AA v2.1 and 45 on AA-Briefcase v1.1 - Highest score on the Artificial Analysis Intelligence Index: 38 (Astra: 53) One safety number worth knowing, from OpenAI's [[GPT-6.1 Sol]] post: when a search tool is broken, GPT-6 Luna fails to tell the user in 28.7% of the test cases (GPT-6 Sol: 4.9%, Astra: 1.5%). These tasks are built to provoke failures and don't reflect typical usage, but if you use Luna as an unattended agent, check its tool results yourself. ## Where it fits OpenAI's own model-selection guide puts Luna at the bottom of the ladder: - **Luna at low effort**: fine-grained edits, well-scoped problem solving, simple data extraction - **Luna at `xhigh`**: finding context across several apps, prioritizing work, solving problems with clear constraints - Anything more complex goes to Sol, then Astra Practitioners use it the same way. Simon Willison calls Luna his favorite model for building product features because of its cost and speed, and moved the Datasette Agent demo to it on launch day. Developers on [[Hacker News]] described setups with Sol as planner and reviewer and Luna at max as the coder subagent, or Luna at `xhigh` as a replacement for the discontinued Terra. At DevDay, OpenAI previewed a Decisions API that uses Luna to pick from a predefined set of options "in a fraction of a second". ## My take This release is a price cut. Benchmarks moved by a couple of points in both directions; cost per task dropped by more than half. Luna is still the most important model in the family for most builders. The stage-splitting pattern from [[GPT-5.6]] gets cheaper again: let a bigger model plan, let Luna do the well-specified work, and measure where quality drops with your own evals. Just don't hand it ambiguous tasks or unchecked tools; the broken-tool disclosure number above is the reason. ## References - OpenAI, "Introducing GPT-6 Sol and Luna" (2026-09-22): https://openai.com/index/introducing-gpt-6-sol-and-luna/ - OpenAI launch post on X: https://x.com/OpenAI/status/2102460975790137662 - OpenAI API model docs: https://developers.openai.com/api/docs/models/gpt-6-luna - OpenAI, model selection guide: https://developers.openai.com/api/docs/guides/model-selection - OpenAI, "Introducing GPT-6.1 Sol" (broken search tool disclosure eval): https://openai.com/index/introducing-gpt-6-1-sol/ - Artificial Analysis on X (benchmark results): https://x.com/ArtificialAnlys/status/2102462962758033624 - Artificial Analysis, GPT-6 Luna model page: https://artificialanalysis.ai/models/releases/gpt-6-luna - Simon Willison on X: https://x.com/simonw/status/2102490016752779458 - Simon Willison, "Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war": https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/ - Simon Willison, OpenAI DevDay 2026 live blog: https://simonwillison.net/2026/Sep/29/openai-devday-2026-live-blog/ - Hacker News, "GPT-6 Sol and Luna": https://news.ycombinator.com/item?id=49805509 ## Related - [[GPT-6 Sol]] - [[GPT-6.1 Sol]] - [[GPT-6 Astra]] - [[2026-09-22 GPT-6 Sol and Luna - same smarts as GPT-5.6, half the price]] - [[GPT-5.6]] - [[OpenAI]] - [[OpenAI Codex]] - [[ChatGPT]] - [[Claude Opus 5]] - [[Claude Fable 5]] - [[Artificial Analysis]] - [[Large Language Models (LLMs)]] - [[Context Window]] - [[AI Agents]] - [[Ori Eval]] - [[Simon Willison]] - [[Hacker News]]