GPT-6 Luna
GPT-6 Luna is the cheapest and fastest model in the GPT-6 family, released by OpenAI on Sep 22, 2026. It costs US$ 0.10 per million input tokens and US$ 0.50 per million output, rereads the cache for US$ 0.01 and works with 1.05 million tokens of context. It suits clear, repeatable tasks such as extraction, classification and structured summaries; for open-ended, complex work, OpenAI itself points to GPT-6.1 Sol or Astra.
GPT-6 Luna in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- OpenAI
- API identifier
gpt-6-luna- Release
- Sep 22, 2026, together with GPT-6 Sol
- Input
- US$ 0.10 per million tokens
- Output
- US$ 0.50 per million tokens
- Cache read
- US$ 0.01 per million (10% of input)
- Cache write
- US$ 0.125 per million (1.25x input)
- Above 272 thousand input tokens
- US$ 0.20 input and US$ 0.75 output, for the whole request
- Batch and Flex
- US$ 0.05 input and US$ 0.25 output
- Fast mode
- US$ 0.20 and US$ 1
- Context window
- 1,050,000 tokens (up to 922 thousand input)
- Max output
- 128 thousand tokens
- Knowledge cutoff
- May 18, 2026, the most recent in the family
- Reasoning
- effort
nonetomax, defaultmedium - API limit at tier 5
- 30 thousand requests and 180 million tokens per minute
- Where it runs
- OpenAI API, ChatGPT Work, Codex, the desktop app on the Free and Go plans, and Amazon Bedrock (us-east-1)
- Cyber risk
- High on the Preparedness scale
Where GPT-6 Luna does well, where it falls behind and who it is for
Where it does well
- Price: US$ 0.10 and US$ 0.50 per million tokens, one tenth of Claude Haiku 4.5 (US$ 1 and US$ 5) and one twentieth of GPT-6 Sol.
- Coding for the money: 66.6% on DeepSWE at max effort, for US$ 0.22 per task. OpenAI says this is on par with Claude Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost per task.
- Computer use: 52.7% on OSWorld 2.0 offline at max effort, above GPT-5.6 Sol at medium effort, at one tenth of the cost per task.
- Volume: at API tier 5, 30 thousand requests and 180 million tokens per minute, against 15 thousand and 40 million for Sol and Astra.
- Window and freshness: 1.05 million tokens, the same as Astra, and a knowledge cutoff of May 18, 2026, the most recent of the four GPT-6 models.
Where it falls behind
- Low effort: at low effort it scored 2.4% on DeepSWE and 1.2% on AutomationBench. In Codex, OpenAI recommends starting at high effort.
- Complex automation: 20.7% on AutomationBench at its best setting, against 33.2% for GPT-6 Sol and 41.4% for Astra.
- Refusals: OpenAI says some of Luna's high jailbreak scores may reflect a broader tendency to refuse requests, including legitimate ones.
- Images: until Sep 25, 2026, an encoding bug degraded image understanding. OpenAI fixed it and recommends rerunning evaluations with images.
- Tools on Chat Completions: only at
noneeffort. To call functions with reasoning, you need the Responses API.
Who it makes sense for
- High-volume extraction, classification, transformation and structured summaries, the list OpenAI itself gives for Luna.
- Targeted edits to text or code and simple data extraction, with results that are easy to check.
- The first filter in a pipeline: Luna handles the simple cases and passes the hard ones to GPT-6.1 Sol or Astra.
- Long documents read in full at low cost: even above 272 thousand tokens, input stays at US$ 0.20 per million.
When to pick another model
- Open-ended, complex tasks without a clear success criterion: OpenAI points to GPT-6.1 Sol or Astra.
- Automation with many steps and tools, where GPT-6 Sol scores 33.2% against 20.7% on AutomationBench.
- Coding or agent work at low effort: scores collapse, and at high effort the cost per task is still a few cents.
- Customer service where refusing a legitimate request is costly, without first testing on real cases.
What OpenAI published at launch
Luna came out on Sep 22, 2026 with GPT-6 Sol. It succeeds GPT-5.6 Luna, which OpenAI compares to the old nano tier of the GPT-5 family. Input was halved from its predecessor's promotional price, from US$ 0.20 to US$ 0.10, and output fell further, from US$ 1.20 to US$ 0.50.
The announcement charts show two things. First, the cost per task stays under US$ 0.30 on every test, even at max effort. Second, the model depends heavily on effort: from low to max, its DeepSWE score goes from 2.4% to 66.6%. On factuality, OpenAI says Luna at higher effort levels matches GPT-5.6 Sol at about a hundredth of its cost.
| GPT-6 Luna OpenAI | GPT-5.6 Luna OpenAI | GPT-6 Sol OpenAI | Claude Opus 5 Anthropic | |
|---|---|---|---|---|
| AutomationBench automation |
20.7% (US$ 0.037) | 17.0% (US$ 0.07) | 33.2% (US$ 0.27) | 26.9% (US$ 3.05) |
| Agents' Last Exam agent tasks |
50.9% (US$ 0.15) | 50.4% (US$ 2.57) | 56.4% (US$ 2.93) | 55.9% (US$ 7.29) |
| DeepSWE v1.1 software engineering |
66.6% (US$ 0.22) | 62.2% (US$ 0.53) | 68.8% (US$ 2.74) | 73.7% (US$ 11.84) |
| FrontierCode 1.1 Main coding |
42.4% (US$ 0.11) | 39.8% (US$ 0.37) | 49.3% (US$ 2.14) | 53.4% (US$ 4.31) |
| OSWorld 2.0 offline computer use |
52.7% (US$ 0.27) | 52.7% (US$ 0.49) | 64.4% (US$ 3.25) | 70.2% (US$ 24.11) |
| Factual errors hard prompts, lower is better |
7.6% (US$ 0.012) | 12.0% (US$ 0.022) | 4.5% (US$ 0.13) | not measured |
Charts from the GPT-6 Sol and Luna announcement. Each model's best score at any effort level; in brackets, the average cost per task. The dot marks the best in each row. On OSWorld, Luna ties its predecessor on score and costs a little over half as much per task.
Luna against Claude Haiku 4.5: price
At Anthropic, the closest model on price is Claude Haiku 4.5, at US$ 1 input and US$ 5 output per million tokens. Luna costs one tenth on both lines and rereads the cache for US$ 0.01, against US$ 0.10. Each maker calls its model the fastest in its lineup.
The window matters as much as the price. Haiku 4.5 works with 200 thousand tokens; Luna, with 1.05 million. Above 272 thousand input tokens Luna doubles its input price, but a request that size does not even fit in Haiku 4.5. No maker table used here compares the quality of the two: a test with your own tasks decides.
| GPT-6 Luna OpenAI | Haiku 4.5 Anthropic | GPT-6 Sol OpenAI | |
|---|---|---|---|
| Input | US$ 0.10 | US$ 1 | US$ 2 |
| Output | US$ 0.50 | US$ 5 | US$ 10 |
| Cache read | US$ 0.01 | US$ 0.10 | US$ 0.20 |
| Input above 272 thousand tokens | US$ 0.20 | does not fit the window | US$ 4 |
| Context window | 1.05 million | 200 thousand | 1.05 million |
OpenAI and Anthropic pricing pages, checked on Oct 1, 2026.
Astra, 6.1 Sol, Sol or Luna: choosing within OpenAI
Luna is the base of the family. OpenAI describes it as the fastest and most cost-effective, strong at focused, high-volume tasks. In Codex, the advice is to choose Luna when you know what a good result looks like, and to start at high effort.
Moving up is expensive: GPT-6.1 Sol costs 20 times as much per token, and Astra 100 times. That is why a tiered pipeline makes sense, with Luna handling what is clear and repetitive and passing on only what needs more.
| GPT-6 Astra | GPT-6.1 Sol | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|---|
| Input | US$ 10 | US$ 2 | US$ 2 | US$ 0.10 |
| Output | US$ 50 | US$ 10 | US$ 10 | US$ 0.50 |
| Cache read | US$ 1 | US$ 0.10 | US$ 0.20 | US$ 0.01 |
| Cache write | US$ 12.50 | US$ 2.50 | US$ 2.50 | US$ 0.125 |
| Above 272 thousand tokens input and output |
US$ 20 and US$ 75 | US$ 4 and US$ 15 | US$ 4 and US$ 15 | US$ 0.20 and US$ 0.75 |
| Context window | 1.05 million | 1.05 million | 1.05 million | 1.05 million |
| Knowledge cutoff | Apr 30, 2026 | Apr 30, 2026 | Apr 20, 2026 | May 18, 2026 |
| Accepts effort none | no | no | yes | yes |
| Role, per OpenAI | highest intelligence | balance of cost and intelligence | superseded by 6.1 Sol | fastest and most cost-effective |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026.
Examples of using GPT-6 Luna
Every incoming message is classified by topic, urgency and product, and only the ambiguous cases go up to a larger model. At US$ 0.10 per million input tokens, volume stops being the problem on the bill.
Invoices, contracts and forms become structured fields for your system. It is the kind of task OpenAI lists for Luna, with results that are easy to check against the source.
A main agent splits the work and hands Luna targeted fixes and simple tests. On DeepSWE, Luna at max effort scored 66.6% for US$ 0.22 per task, in OpenAI's figures.
Where GPT-6 Luna stands out and where it falls behind, by the numbers
GPT-6 Luna has published results on 23 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 26, 2025 and Dec 21, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 23 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 90.5% | #35 / 73 | #37 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 98.9% | #17 / 75 (tied with 3) | #17 / 168 | 100.0% GPT-5.5 |
| Dtbench | 90.1% | #36 / 64 (tied with 2) | #38 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 44.49 | #35 / 64 | #35 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 19.4% | #27 / 76 | #27 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 54.6% | #29 / 74 | #29 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 78.9% | #18 / 64 | #18 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 59.3% | #32 / 43 | #32 / 83 | 95.0% GPT-6 Astra |
| WebDev Arena | 1,582 | #23 / 67 | #23 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 86.7% | #33 / 43 | #34 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 41.4% | #37 / 59 | #43 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 36.0% | #33 / 58 (tied with 1) | #39 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 7.0% | #56 / 58 (tied with 1) | #69 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 64.0% | #18 / 64 | #18 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 56.1% | #17 / 54 (tied with 1) | #17 / 67 | 100.0% GPT-6.1 Sol |
| Apex Agents | 44.3% | #34 / 48 | #34 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 42.4% | #22 / 43 | #22 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 23.0% | #19 / 36 | #19 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 44.2% | #13 / 32 (tied with 1) | #13 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 31.2% | #14 / 29 (tied with 1) | #14 / 31 | 54.4% Gemini 4 Argon |
| Spireclimb | 0.0% | #7 / 7 | #7 / 7 | 25.7% GPT-6 Astra |
| Spireclimb Sts1 A20 Heart | 0.0% | #4 / 7 (tied with 3) | #4 / 7 | 20.0% GPT-6.1 Sol |
| Spireclimb Sts2 A10 | 0.0% | #7 / 7 | #7 / 7 | 34.7% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare GPT-6 Luna with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
| Claude Opus 4.7 | Anthropic | 156.32 | 1,532 | 76.5 | Apr 16, 2026 |
| Grok 4.6 | xAI | 156.31 | 1,475 | 78.0 | Aug 12, 2026 |
| GPT-6 Luna | OpenAI | 156.31 | n/a | 72.0 | Sep 22, 2026 |
| GLM-5.3 | Z.ai (Zhipu AI) | 155.91 | n/a | 76.1 | Aug 14, 2026 |
GPT-6 Luna or Grok 4.6?
Grok 4.6 is the model from another lab closest to GPT-6 Luna in capability. Across the 20 tests both took, this is the score:
| Test | GPT-6 Luna | Grok 4.6 |
|---|---|---|
| Game puzzles | 7.0% | 34.0% |
| Frontiermath Tier 4 V2 | 56.1% | 31.7% |
| Apex Agents | 44.3% | 65.3% |
| Gdp Pdf | 23.0% | 17.2% |
| ProofBench | 64.0% | 51.0% |
| Frontiermath Tiers 1 3 V2 | 78.9% | 66.0% |
| SimpleQA Verified | 41.4% | 49.3% |
| FrontierCode | 42.4% | 48.0% |
| ARC-AGI 2 | 59.3% | 67.1% |
| Furniture Assembly | 44.2% | 40.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
Other models from OpenAI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| GPT-6.1 Sol | Sep 29, 2026 | 166.32 | n/a | 81.6 | API access |
| GPT-6 Sol | Sep 22, 2026 | 162.55 | n/a | 79.3 | API access |
| GPT-6 Luna | Sep 22, 2026 | 156.31 | n/a | 72.0 | API access |
| GPT-6 Astra | Sep 3, 2026 | 166.41 | n/a | 82.2 | API access |
| GPT-5.6 Sol | Jul 9, 2026 | 161.73 | 1,527 | 81.1 | API access |
| GPT-5.6 Terra | Jul 9, 2026 | 159.43 | 1,515 | 77.9 | API access |
| GPT-5.6 Luna | Jul 9, 2026 | 156.33 | 1,478 | 73.6 | API access |
| GPT-5.5 Instant | May 5, 2026 | 142.45 | 1,456 | n/a | API access |
| GPT-5.5 Pro | Apr 23, 2026 | 161.97 | n/a | n/a | API access |
Have you used GPT-6 Luna? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about GPT-6 Luna
How much does GPT-6 Luna cost?
US$ 0.10 per million input tokens and US$ 0.50 per million output. Cache reads cost US$ 0.01 and cache writes US$ 0.125. Above 272 thousand input tokens, the whole request moves to US$ 0.20 input and US$ 0.75 output. Batch and Flex are half price.
GPT-6 Luna or Claude Haiku 4.5?
On price, Luna costs one tenth of Haiku 4.5 for input and output, and its window is 1.05 million tokens against 200 thousand. No maker table used here compares the quality of the two. The independent scores are at the top of this page, and a test with your own tasks decides.
Is GPT-6 Luna good for coding?
For focused tasks, yes: it scored 66.6% on DeepSWE at max effort for US$ 0.22 per task, in OpenAI's figures. At low effort it dropped to 2.4%. For open-ended, complex coding work, OpenAI points to GPT-6.1 Sol or Astra.
Can I use GPT-6 Luna for free?
In ChatGPT, the Free and Go plans get GPT-6 Luna in the desktop app. The Plus, Pro, Business, Enterprise and Edu plans get Luna in ChatGPT Work and Codex. On the API, usage is billed per token.
Where does GPT-6 Luna rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-6 Luna ranks #27 out of 267 models, with an ECI of 156.31. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is GPT-6 Luna better than Grok 4.6?
Across the 20 tests both models took and Epoch AI compiled, GPT-6 Luna comes out ahead on 6, behind on 14 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is GPT-6 Luna best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of GPT-6 Luna are: AIME 2024 and 2025 (OTIS) (#17 of 75); Frontiermath Tiers 1 3 V2 (#18 of 64); ProofBench (#18 of 64). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- OpenAI, GPT-6 Sol and GPT-6 Luna announcement
- OpenAI, GPT-6 Luna model page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6 Astra system card (GPT-6 Sol and GPT-6 Luna appendix)
- OpenAI, GPT-6.1 Sol addendum (note on Luna refusals)
- OpenAI, API changelog (image fix on Sep 25, 2026)
- OpenAI, latest model guide
- OpenAI, model selection guide
- OpenAI, models in Codex and ChatGPT Work
- OpenAI, GPT-5.6 Luna model page (nano tier)
- OpenAI, models on Amazon Bedrock (documentation)
- Anthropic, pricing (documentation)
- Anthropic, models overview (documentation)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking