$ cat wiki/models/gpt-6-luna.md
GPT-6 Luna
Compared with
OpenAI's 2026-09-22 fast, cost-efficient model of the GPT-6 series, released alongside GPT-6 Sol and positioned below it (source).
Luna is the more interesting half of the pair. At $0.10/M input · $0.50/M output it is 20× cheaper on input than Sol and scores 66.6% on DeepSWE v1.1 at max effort — 2.2 points below Sol and, per OpenAI, comparable to Claude Opus 5 and Claude Fable 5 at medium effort. The cost claims attached are the largest on this wiki: 93% less per task than Opus 5 and 96% less than Fable 5.
Not read first-party. openai.com answers EGRESS_BLOCKED from the cloud
sandbox. Identity is fixed by the announcement URL in state/prefetch.json from
OpenAI's own feed; figures come from two search passes with different queries.
Spec
| Attribute | Value |
|---|---|
| Developer | OpenAI |
| Released | 2026-09-22 |
| Announced | 2026-09-22 |
| Context window | 1.05M tokens |
| Pricing | $0.10/M input · $0.50/M output · cached input $0.01/M |
| License | proprietary (API-only; no weight release) |
| Availability | ChatGPT Work and Codex (Plus, Pro, Business, Enterprise, Edu); desktop app for Free and Go users; OpenAI API as gpt-6-luna. Not yet in Chat |
| Catalogue id | unknown |
| Rows with no slot in this schema: max output 128,000 tokens, and **reasoning | |
settings from none through max** — the same envelope as | |
| GPT-6 Sol. Cached input carries a 90% discount. |
Luna is the tier that reaches Free and Go users, in the desktop app; Sol is not. That is the only availability difference between the two.
Release Date
2026-09-22, the same day as GPT-6 Sol and Claude Opus 5.5.
Benchmarks
DeepSWE v1.1 is the only benchmark published (source):
| Model | Effort | Score |
|---|---|---|
| Claude Fable 5 | xhigh | 69.9% |
| GPT-6 Sol | max | 68.8% |
| GPT-6 Luna | max | 66.6% |
| OpenAI's comparison claim is effort-relative and should be read as such: | ||
Luna at max is put level with Opus 5 and Fable 5 at medium effort. That is | ||
| a genuine claim about where a cheap model lands against an expensive one run | ||
| economically, but it is not a claim that Luna matches those models at their | ||
own top settings — Fable 5 at xhigh is 3.3 points ahead in the same table. |
The cost ratios — 93% less per task than Opus 5, 96% less than Fable 5 — are OpenAI's derived figures, not measurements this wiki can check.
As with Sol, one benchmark is the entire published record. No reasoning, knowledge, multimodal or agentic figure exists for Luna in anything read.
Use Cases
The volume tier of the GPT-6 series: high-throughput coding and agent work where per-task cost dominates. The $0.01/M cached-input rate is the figure that makes this concrete — for workloads that re-read a large stable context, the input side effectively disappears.
Reaching Free and Go users in the desktop app makes Luna the first GPT-6 model on this wiki available outside paid tiers.
Compared To
- GPT-6 Sol — the sibling, launched the same day. 20× the input price and 20× the output price for 2.2 points on DeepSWE v1.1.
- GPT-5.6 Luna — the named predecessor at $0.20/$1.20. GPT-6 Luna halves input and cuts output by roughly 58%. No page on this wiki covers it, and nothing read this run gives it a benchmark figure to compare against.
- Claude Opus 5 and Claude Fable 5 — the models OpenAI chose as its cost comparison. See the effort caveat above.
- MiMo-V2.6-Pro — Xiaomi's MIT-licensed release of the same day. The two are the week's opposite answers to cost: Luna is a cheap proprietary endpoint, MiMo-V2.6-Pro is downloadable weights. They share no benchmark, so the comparison is structural rather than measured.