GPT-5.6 Luna
GPT-5.6 Luna is the smallest and cheapest model in the GPT-5.6 family, released by OpenAI on Jul 9, 2026, roughly equivalent to the old nano tier. It costs US$ 0.20 per million input tokens and US$ 1.20 per million output since the 80% price cut of Jul 30, with a 1.05 million token window. On Sep 22 it got a successor, GPT-6 Luna, at half the input price and less than half the output price. Even so, it remains the default ChatGPT model on the Free and Go plans and has no shutdown date on the API.
GPT-5.6 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-5.6-luna- Release
- Jul 9, 2026, with GPT-5.6 Sol and GPT-5.6 Terra
- Input
- US$ 0.20 per million tokens (US$ 1 at launch)
- Output
- US$ 1.20 per million tokens (US$ 6 at launch)
- Cache read
- US$ 0.02 per million (10% of input)
- Cache write
- US$ 0.25 per million (1.25x input)
- Above 272 thousand input tokens
- US$ 0.40 input and US$ 1.80 output, for the whole request
- Batch and Flex
- US$ 0.10 input and US$ 0.60 output
- Fast mode
- US$ 0.40 and US$ 2.40
- Context window
- 1,050,000 tokens (up to 922 thousand of input)
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Feb 16, 2026
- Reasoning
- effort
nonetomax, defaultmedium - API Tier 5 limit
- 30 thousand requests and 180 million tokens per minute
- Where it runs
- OpenAI API, ChatGPT (default on Free and Go plans), ChatGPT Work and Codex
- Cyber risk
- High on the Preparedness scale, like Sol and Terra
- Status
- successor GPT-6 Luna since Sep 22, 2026; no deprecation notice on the API
Where GPT-5.6 Luna does well, where it falls behind and who it is for
Where it does well
- Price: US$ 0.20 and US$ 1.20 per million tokens, one fifth of the launch price and one tenth of GPT-5.6 Terra.
- Agent tasks: 50.3% on Agents' Last Exam, above GPT-5.5 (46.9%), Claude Opus 4.8 (45.2%) and Claude Fable 5 (40.5%), in OpenAI's table.
- Coding: 62.7% on SWE-Bench Pro and 74.6 on the Artificial Analysis coding agent index, where OpenAI says it beats Claude Opus 4.8 (72.5).
- Window and volume: a 1.05 million token context and, at API Tier 5, 30 thousand requests and 180 million tokens per minute.
- Shared reasoning: it reuses the saved reasoning of GPT-5.6 Sol and Terra in the same conversation, which does not happen across different families.
Where it falls behind
- Very long context: 41.3% on MRCR v2 between 256 thousand and 512 thousand tokens, against 89.6% for Terra and 81.5% for GPT-5.5.
- A better, cheaper successor: GPT-6 Luna costs US$ 0.10 and US$ 0.50 and, according to OpenAI, at high effort beats its predecessor by 5.4 points on AutomationBench at 58% lower cost per task.
- Low effort: in OpenAI's chart, Luna's DeepSWE score goes from 1.2% at low effort and 9.3% at medium to 62.2% at max.
- Agents without confirmation: in realistic work tests without a confirmation policy, 19.7% misaligned outcomes, against 3.4% for GPT-6 Astra; with the default policy, 7.2%.
- Knowledge: it runs to Feb 16, 2026, against May 18, 2026 for GPT-6 Luna.
Who it makes sense for
- High volume of simple tasks with reasoning, such as classifying, extracting and summarizing: it is the old nano tier, with the same input price as GPT-5.4 nano (US$ 0.20).
- Integrations that came from
gpt-4.1-nanoorgpt-5-nano-2025-08-07: OpenAI namesgpt-5.6-lunaas the replacement for both, which leave the API on Oct 23, 2026 and Dec 11, 2026. - Data residency in the United Arab Emirates: Luna is on the region's list for Chat Completions and Responses, and no GPT-6 model is.
- Pipelines that mix GPT-5.6 Sol and Luna in the same conversation, to reuse reasoning between the two.
When to pick another model
- New projects: GPT-6 Luna costs half on input, less than half on output and has more recent knowledge.
- Long documents with scattered information, where Luna's score drops to 41.3% above 256 thousand tokens.
- Agents that handle money or data without asking for confirmation: without the policy, 33.5% unauthorized transactions in the Astra system card tests.
- Open-ended, complex coding: within the family, Sol scored higher than Luna on every coding test in the July table.
What OpenAI published at launch
OpenAI released GPT-5.6 in three tiers: Sol, the flagship; Terra, the middle option; and Luna, the fastest and cheapest. The Luna page says it roughly corresponds to the nano tier of earlier GPT-5 families. The launch price was US$ 1 and US$ 6; three weeks later, on Jul 30, it fell 80%, to US$ 0.20 and US$ 1.20.
In the announcement table, Luna leads none of the ten tests below, but it stays close to Terra, which now costs ten times more, and beats GPT-5.5, which costs US$ 5 and US$ 30, on three of them. Its biggest weakness is very long context.
| GPT-5.6 Luna OpenAI | GPT-5.6 Terra OpenAI | GPT-5.5 OpenAI | Claude Opus 4.8 Anthropic | |
|---|---|---|---|---|
| Agents' Last Exam long professional tasks |
50.3% | 50.4% | 46.9% | 45.2% |
| GDPval-AA v2 office work, Elo |
1,591.8 | 1,593 | 1,493.7 | 1,600.1 |
| AA Coding Agent Index v1.1 coding agent |
74.6 | 77.4 | 76.4 | 72.5 |
| SWE-Bench Pro software engineering |
62.7% | 63.4% | 59.4% | 69.2% |
| Terminal-Bench 2.1 terminal work |
84.7% | 87.4% | 85.6% | 78.9% |
| OSWorld 2.0 computer use |
45.6% | 50.2% | 47.5% | 54.8% |
| BrowseComp web research |
83.3% | 87.5% | 84.4% | 84.3% |
| GPQA Diamond PhD-level science |
92.3% | 92.9% | 93.6% | 92% |
| Toolathlon tool use |
53.4% | 53.1% | 55.6% | 59.9% |
| MRCR v2, 256K to 512K long context |
41.3% | 89.6% | 81.5% | not measured |
Table from the GPT-5.6 announcement (Jul 9, 2026); Claude figures were published by OpenAI. The dot marks the best in each row. "Not measured" means no published figure. According to OpenAI, Agents' Last Exam measures long professional workflows across 55 fields.
Luna against its successor, on the same chart
In the GPT-6 Luna announcement on Sep 22, OpenAI published charts with both models side by side, showing the score and average cost per task at each effort level. GPT-6 Luna scores higher on five of the six tests, ties on computer use and costs less per task on all of them.
The GPT-5.6 Luna numbers in those charts do not match the July table: on DeepSWE, the July table showed 67.2% and the September chart shows 62.2%. That is why the comparison below uses only the September chart, where both appear together.
| GPT-5.6 Luna Jul 2026 | GPT-6 Luna Sep 2026 | |
|---|---|---|
| AutomationBench automation |
17.0% (US$ 0.070) | 20.7% (US$ 0.037) |
| Agents' Last Exam agent tasks |
50.4% (US$ 2.57) | 50.9% (US$ 0.15) |
| FrontierCode 1.1 Main coding |
39.8% (US$ 0.37) | 42.4% (US$ 0.11) |
| DeepSWE v1.1 software engineering |
62.2% (US$ 0.53) | 66.6% (US$ 0.22) |
| OSWorld 2.0 offline computer use |
52.7% (US$ 0.49) | 52.7% (US$ 0.27) |
| Factual errors hard prompts, lower is better |
12.0% (US$ 0.022) | 7.6% (US$ 0.012) |
| Price per million tokens input and output |
US$ 0.20 and US$ 1.20 | US$ 0.10 and US$ 0.50 |
Charts from the GPT-6 Sol and GPT-6 Luna announcement (Sep 22, 2026). Best score of each model, all at max effort; in parentheses, the average cost per task. The dot marks the best in each row; OSWorld is a tie on score.
Is GPT-5.6 Luna still worth using in 2026?
For a new API project, no. GPT-6 Luna costs half on input and 58% less on output, has knowledge three months more recent and, in OpenAI's charts, scores higher while spending less per task.
GPT-5.6 Luna is still in use, though. It is the default ChatGPT model on the Free and Go plans and also answers the Think button on those plans; GPT-6 Luna has not reached chat yet, only ChatGPT Work, Codex and the desktop app. On the API there is no shutdown date, and it is the official replacement for gpt-4.1-nano and gpt-5-nano-2025-08-07.
| GPT-5.6 Luna Jul 2026 | GPT-6 Luna Sep 2026 | GPT-6.1 Sol Sep 2026 | GPT-6 Astra Sep 2026 | |
|---|---|---|---|---|
| Input | US$ 0.20 | US$ 0.10 | US$ 2 | US$ 10 |
| Output | US$ 1.20 | US$ 0.50 | US$ 10 | US$ 50 |
| Cache read | US$ 0.02 | US$ 0.01 | US$ 0.10 | US$ 1 |
| Cache write | US$ 0.25 | US$ 0.125 | US$ 2.50 | US$ 12.50 |
| Above 272 thousand tokens input and output |
US$ 0.40 and US$ 1.80 | US$ 0.20 and US$ 0.75 | US$ 4 and US$ 15 | US$ 20 and US$ 75 |
| Knowledge cutoff | Feb 16, 2026 | May 18, 2026 | Apr 30, 2026 | Apr 30, 2026 |
| Accepts none effort | yes | yes | no | no |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026.
What changes when replacing GPT-5.6 Luna with GPT-6 Luna
The switch touches little: the window, effort levels and cache rules stay the same. The points to watch:
- Identifier:
gpt-6-lunainstead ofgpt-5.6-luna. - Tools on Chat Completions: on GPT-6 Luna, function calling there only works with
noneeffort. With reasoning, use the Responses API. - Saved reasoning: OpenAI only reuses reasoning within the same model family, so reasoning from GPT-5.6 conversations does not carry over to GPT-6 Luna.
- Data residency: GPT-6 Luna is not on the United Arab Emirates list, and its Fast mode does not work with European Union data residency.
- Images: an encoding bug hurt GPT-6 Luna's image understanding until Sep 25, 2026. OpenAI fixed it and recommends rerunning image evaluations.
- Knowledge: from Feb 16, 2026 to May 18, 2026.
Examples of using GPT-5.6 Luna
Every incoming message is classified by topic and urgency, and only ambiguous cases go up to a larger model. At US$ 0.20 per million input tokens, this is the kind of cost-sensitive, high-volume workload OpenAI designed Luna for.
Invoices, contracts and forms become structured fields for the system. Luna supports structured outputs and reads up to 922 thousand input tokens in one call.
An agent uses GPT-5.6 Sol to plan and Luna to run the simple steps. Since both belong to the same family, saved reasoning passes from one to the other.
Where GPT-5.6 Luna stands out and where it falls behind, by the numbers
GPT-5.6 Luna has published results on 27 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Oct 12, 2025 and Oct 7, 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 27 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 91.6% | #26 / 81 | #26 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 98.3% | #21 / 82 (tied with 3) | #22 / 168 | 100.0% GPT-5.5 |
| Dtbench | 88.8% | #45 / 70 | #46 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 48.54 | #24 / 70 | #24 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 20.6% | #23 / 79 | #24 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 60.9% | #28 / 55 | #28 / 112 | 93.6% GPT-6 Astra |
| SciCode | 53.6% | #33 / 76 | #34 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 82.1% | #15 / 66 | #15 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 59.5% | #31 / 51 | #31 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 1,667 | #11 / 65 | #11 / 82 | 2,951 GPT-6 Astra |
| ARC-AGI | 88.0% | #30 / 51 | #30 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 41.0% | #40 / 64 (tied with 1) | #45 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 43.0% | #28 / 65 (tied with 1) | #29 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 21.0% | #38 / 61 | #41 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 60.0% | #19 / 69 | #20 / 74 | 100.0% Claude Fable 5.1 |
| SimpleBench | 46.8% | #32 / 38 (tied with 1) | #40 / 74 | 81.9% Claude Fable 5 |
| Frontiermath Tier 4 V2 | 61.0% | #14 / 56 | #15 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 4,095 | #35 / 49 | #36 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 43.0% | #35 / 49 (tied with 1) | #35 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 39.8% | #26 / 43 | #26 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 22.7% | #20 / 36 | #21 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 42.5% | #14 / 33 (tied with 1) | #15 / 34 | 83.3% Claude Opus 5.5 |
| BALROG | 45.6% | #8 / 10 | #8 / 32 | 68.3% GPT-6 Astra |
| Blueprint Bench 2 | 22.6% | #24 / 31 | #24 / 31 | 54.4% Gemini 4 Argon |
| DeepSWE | 67.2% | #10 / 26 | #10 / 26 | 74.1% GPT-6 Astra |
| Surface Evolver Bench | 61.9% | #9 / 23 | #9 / 24 | 95.0% Claude Fable 5 |
| CursorBench | 35.9% | #15 / 16 | #15 / 16 | 57.8% Claude Opus 5.5 |
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-5.6 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 |
| GPT-5.6 Luna | OpenAI | 156.33 | 1,478 | 73.6 | Jul 9, 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 |
| GLM-5.3 | Z.ai (Zhipu AI) | 155.91 | n/a | 76.1 | Aug 14, 2026 |
GPT-5.6 Luna or Claude Opus 4.7?
Claude Opus 4.7 is the model from another lab closest to GPT-5.6 Luna in capability. Across the 23 tests both took, this is the score:
| Test | GPT-5.6 Luna | Claude Opus 4.7 |
|---|---|---|
| Vending Bench 2 | 4,095 | 10,937 |
| Frontiermath Tier 4 V2 | 61.0% | 31.7% |
| CritPt | 20.6% | 12.0% |
| Game puzzles | 21.0% | 28.0% |
| SimpleBench | 46.8% | 61.7% |
| Furniture Assembly | 42.5% | 33.3% |
| ARC-AGI 2 | 59.5% | 75.8% |
| SimpleQA Verified | 41.0% | 51.7% |
| ALE-Bench | 1,667 | 1,323 |
| WeirdML | 60.9% | 76.4% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What GPT-5.6 Luna costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
ARC-AGI 2
The score scale does not start at zero.
ARC-AGI
The score scale does not start at zero.
CursorBench
The score scale does not start at zero.
Other models from OpenAI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | 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 |
| GPT-5.5 | Apr 23, 2026 | 158.95 | 1,533 | 80.2 | API access |
Have you used GPT-5.6 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-5.6 Luna
How much does GPT-5.6 Luna cost?
US$ 0.20 per million input tokens and US$ 1.20 per million output. Cache reads cost US$ 0.02 and cache writes US$ 0.25. Above 272 thousand input tokens, the whole request moves to US$ 0.40 and US$ 1.80. Batch and Flex are half price, and Fast mode costs double.
Did the price of GPT-5.6 Luna change?
Yes. Luna launched at US$ 1 and US$ 6 on Jul 9, 2026, and OpenAI cut the price by 80% starting Jul 30, to US$ 0.20 and US$ 1.20. On the same day, GPT-5.6 Terra became 20% cheaper.
GPT-5.6 Luna or GPT-6 Luna?
For a new project, GPT-6 Luna: it costs US$ 0.10 and US$ 0.50 and, in OpenAI's charts, scores higher on five of six tests, with a lower cost per task on all of them. 5.6 Luna makes sense where it is already integrated or where data residency in the United Arab Emirates is required.
Does free ChatGPT use GPT-5.6 Luna?
Yes. GPT-5.6 Luna is the default model on the Free and Go plans in ChatGPT and also answers when the user presses Think. GPT-6 Luna is, for now, in ChatGPT Work, Codex and the desktop app, but not in chat.
Is GPT-5.6 Luna being shut down?
There is no published date. OpenAI's deprecations page has no entry for it, and the model is still the named replacement for models that leave the API in October and December 2026. Generally available models get at least six months of notice. The independent scores are at the top of this page.
Where does GPT-5.6 Luna rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-5.6 Luna ranks #25 out of 267 models, with an ECI of 156.33. 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-5.6 Luna better than Claude Opus 4.7?
Across the 23 tests both models took and Epoch AI compiled, GPT-5.6 Luna comes out ahead on 11, behind on 12 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-5.6 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-5.6 Luna are: ALE-Bench (#11 of 65); Frontiermath Tiers 1 3 V2 (#15 of 66); Frontiermath Tier 4 V2 (#14 of 56). 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-5.6 announcement
- OpenAI, GPT-5.6 Luna model page (documentation)
- OpenAI, API pricing
- OpenAI, API changelog (price cut on Jul 30, 2026)
- OpenAI, GPT-5.6 system card
- OpenAI, GPT-6 Sol and GPT-6 Luna announcement
- OpenAI, GPT-6 Astra system card (realistic work tests)
- OpenAI, GPT-6 Luna model page (documentation)
- OpenAI, GPT-5.6 and GPT-6 Pro in ChatGPT (help center)
- OpenAI, API deprecations
- OpenAI, reasoning guide (saved reasoning)
- OpenAI, data residency guide
- OpenAI, latest model guide
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