GPT-6 Luna is OpenAI’s fastest and cheapest GPT-6 model, released on September 22, 2026. It reads text and images, returns text, and supports a 1.05-million-token context window. API prices are $0.10 per million input tokens and $0.50 per million output tokens. OpenAI aims it at high-volume jobs such as extraction, classification and summaries.
GPT-6 Luna is the model OpenAI built for jobs where volume matters more than brilliance. It sits at the bottom of the GPT-6 family, below Sol and the flagship Astra, and it costs a small fraction of either. It is also the GPT-6 model OpenAI has put in front of free ChatGPT users.
This review covers what the GPT-6 Luna AI model can do, what it costs, how it compares with ChatGPT, Sol and last generation’s Luna, and whether it is any good for writing. We built it from OpenAI’s launch material, API listings and independent benchmark reports. We did not run our own tests, so every score below is attributed, and where OpenAI’s numbers and outside testers disagree, we say so.
Last updated: November 10, 2026. OpenAI released four GPT-6 models in four weeks, so check your model picker and the pricing page before you plan around anything here.
Contents
- What GPT-6 Luna is
- Features and specs
- What the benchmarks show
- Luna vs ChatGPT, Sol and GPT-5.6 Luna
- Pricing and free access
- How to use it
- Writing and content creation
- Prompts to try
- Benefits and limitations
- Is it worth using?
- FAQs
- References
What Is GPT-6 Luna?
GPT-6 Luna is a small, fast reasoning model that OpenAI released on September 22, 2026, alongside GPT-6 Sol. Developers call it gpt-6-luna, and plenty of people search for it as “GPT 6 Luna” without the hyphen. It is proprietary, so there are no downloadable weights.
The GPT-6 family arrived in stages. Astra, the flagship, was announced on September 3. Sol and Luna followed on September 22, and GPT-6.1 Sol on September 29, according to Lindy’s GPT-6 breakdown. Luna is the budget end of that line. OpenAI’s own GPT-6 guide points it at focused, repeated work with a clear goal, such as pulling fields from invoices, classifying requests and writing structured summaries.
| Model | Best for | API name | Input / output price |
|---|---|---|---|
| GPT-6 Astra | The hardest work | gpt-6-astra | $10 / $50 |
| GPT-6.1 Sol | Near-Astra results at lower cost | gpt-6.1-sol | $2 / $10 |
| GPT-6 Sol | Complex coding and professional work | gpt-6-sol | $2 / $10 |
| GPT-6 Luna | High-volume, well-defined tasks | gpt-6-luna | $0.10 / $0.50 |
GPT-6 Luna AI Model: Features and Specs
| Spec | Detail |
|---|---|
| Developer | OpenAI (proprietary, no open weights) |
| Released | September 22, 2026 |
| API model ID | gpt-6-luna |
| Input | Text and images (some listings also show files such as PDFs) |
| Output | Text only |
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | May 2026 (listings give May 18, 2026) |
| Reasoning effort | none, low, medium (default), high, xhigh, max |
Reasoning effort is the setting that shapes everything else. Low effort answers fast. Max effort produces the best benchmark scores but thinks for a long time: Artificial Analysis, as reported by NeuralTrust, measured roughly 96 seconds before the first answer token at max. If a person is waiting on the reply, start low.
In the Responses API, Luna supports function calling, web search, computer use and MCP, plus structured outputs and prompt caching. One billing rule catches people out: once a single request passes 272,000 input tokens, the whole request is billed at twice the input rate and 1.5 times the output rate. The million-token window is real, but filling it gets expensive.
OpenAI rates Luna as High capability for cybersecurity and biological risk, so it ships with safeguards. OpenAI also says part of its resistance to jailbreaks comes from a broader tendency to refuse, including some legitimate requests (Unifically summary). If a harmless prompt gets declined, rephrase it before you assume the model is broken.
GPT-6 Luna Capabilities: What the Benchmarks Show
There are two sets of numbers here, and they point in slightly different directions. OpenAI’s results make Luna look like a bargain. Independent testing says it is a bargain mainly because the price fell, not because the model got smarter.
What OpenAI reports
At max effort, OpenAI reports 66.6% on DeepSWE v1.1, a real-world coding benchmark, at about $0.22 per task. That compares with 62.2% for GPT-5.6 Luna. In its launch post, OpenAI says that score is comparable to Claude Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost per task. On Agents’ Last Exam, which runs long professional workflows, Luna scored 50.9% for about $0.15 a task, while GPT-5.6 Luna managed 50.4% for $2.57 (chart data via Unifically).
What independent testers found
Artificial Analysis, which runs its own suite, scores both new models level with their GPT-5.6 predecessors. Luna at max effort cost about $0.07 per task on its index, against $0.18 for GPT-5.6 Luna, but it used more output tokens to get there (about 51,000 against 41,000). Its Coding Agent Index score was 41, two points lower than before.
Knowledge-work tests slipped too. Artificial Analysis recorded a drop of roughly 75 Elo on GDPval-AA, and its team said the regressions tended to come from weaker presentation and from deliverables that skipped parts of the brief. On its hallucination benchmark, Luna improved from a 93% hallucination rate to 77% at max effort, but its summary says Luna got there largely by answering fewer questions, with accuracy about flat at 44%. One hands-on reviewer called it cheap and mostly right, then described a Kubernetes manifest that passed every linter and still crashed on start (Computing for Geeks).
| Measure | Reported by | GPT-6 Luna | For comparison |
|---|---|---|---|
| DeepSWE v1.1 (coding), max effort | OpenAI | 66.6%, about $0.22 per task | GPT-5.6 Luna 62.2%; GPT-6 Sol 68.8% |
| Agents’ Last Exam, max effort | OpenAI | 50.9%, about $0.15 per task | GPT-5.6 Luna 50.4% at $2.57 per task |
| Factual-error rate on OpenAI’s hard prompts (lower is better) | OpenAI | 7.6% at max effort | GPT-5.6 Sol 8.5%; GPT-5.6 Luna 12.0% |
| Cost per task on the Intelligence Index, max effort | Artificial Analysis | About $0.07 | GPT-5.6 Luna about $0.18 |
| Coding Agent Index | Artificial Analysis | 41 | Two points below GPT-5.6 Luna |
| GDPval-AA (knowledge work) | Artificial Analysis | About 75 Elo lower | Versus GPT-5.6 Luna |
| Hallucination rate on AA-Omniscience (lower is better), max effort | Artificial Analysis | 77% | GPT-5.6 Luna 93%; accuracy 44% vs 43% |
The two factuality figures don’t contradict each other. They come from different question sets and different definitions of a mistake. Treat both as a reason to check Luna’s facts, not as a licence to trust them.
GPT-6 Luna vs ChatGPT, Sol and GPT-5.6 Luna
GPT-6 Luna vs ChatGPT
These aren’t rivals. ChatGPT is OpenAI’s app, and Luna is one of the models that runs inside it. At launch, Free and Go users could reach Luna only in the desktop app, while Plus, Pro, Business, Enterprise and Edu users got it in ChatGPT Work and Codex (OpenAI).
That changed within weeks. Notebookcheck, citing OpenAI’s October 7 release notes, reported that GPT-6 began reaching the regular Chat tab for Free and Go users on October 8, with Luna on those plans and Sol on paid ones. The same report noted that OpenAI’s help page still listed GPT-5.6 Luna as the Free and Go default at the time. So when people compare “Luna with ChatGPT”, they usually mean Luna against whatever ChatGPT used to give them, and the answer depends on what your own model picker shows today.
GPT-6 Luna vs GPT-6 Sol
| Factor | GPT-6 Luna | GPT-6 Sol |
|---|---|---|
| API price per million tokens (input / output) | $0.10 / $0.50 | $2 / $10 |
| OpenAI’s positioning | Focused, high-volume tasks | Complex coding, agent and professional work |
| DeepSWE v1.1 at max effort (OpenAI) | 66.6% | 68.8% |
| Intelligence Index at high effort (Artificial Analysis) | 33 | 42 |
| Cost to run that index at high effort | About $48 | About $605 |
| Output speed at high effort | About 134 tokens per second | About 94 tokens per second |
| ChatGPT Chat tab, from the October rollout | Free and Go | Plus, Pro, Business, Enterprise |
Per token, Luna is 20 times cheaper than Sol. Per finished benchmark run it was about 12.6 times cheaper in the NeuralTrust analysis of Artificial Analysis data, because Luna writes more tokens and scores lower. That is still a large saving, but budget with the smaller number.
GPT-6 Luna vs GPT-5.6 Luna
The new Luna costs half as much, and OpenAI’s own tests show similar or better scores. Independent tests show roughly equal intelligence and slightly weaker coding. Read it as a price cut with a model attached. If you already run GPT-5.6 Luna, moving over saves money; don’t expect a jump in quality.
GPT-6 Luna Pricing and Free Access
| Item | Rate |
|---|---|
| Input | $0.10 per million tokens |
| Cached input | $0.01 per million tokens |
| Cache write | $0.125 per million tokens |
| Output | $0.50 per million tokens |
| Batch processing | 50% off standard rates |
| Requests above 272,000 input tokens | 2x input rates and 1.5x output rate for the whole request |
| Previous generation (GPT-5.6 Luna) | $0.20 input, $1.20 output |
OpenAI calls the new prices 50% cheaper, though output fell 58%. Puter’s cost analysis points out that the comparison is against GPT-5.6’s promotional pricing, not a permanent list price.
A worked example, using our own arithmetic: a 1,500-word draft is roughly 2,000 tokens. With a 3,000-token brief going in, one draft costs about $0.0013 on Luna, so a thousand drafts come to about $1.30, or $0.65 in batch mode. The same job on Sol costs about $0.026 per draft. Higher reasoning effort adds tokens, so measure real usage before you extrapolate.
Free access exists, but only inside ChatGPT. Free and Go users can use Luna in the desktop app and, as it rolls out, in the Chat tab. There is no free API tier from OpenAI for it.
How to Use GPT-6 Luna
- In ChatGPT on a Free or Go plan. Open the ChatGPT desktop app, or the Chat tab if the rollout has reached your account. OpenAI called the October rollout global but gave no schedule per account, so you may need to wait.
- In ChatGPT on a paid plan. Plus, Pro, Business, Enterprise and Edu users can pick Luna in Work and Codex. Chat on those plans moved to Sol in October.
- Through the API. Call the Responses API with
gpt-6-lunaand setreasoning.effort. - Through other providers. Luna is listed on OpenRouter, and Lindy reports it on Microsoft Foundry and Amazon Bedrock.
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6-luna",
reasoning={"effort": "low"},
input="Classify this ticket as billing, bug or feature request: My card was charged twice.",
)
print(response.output_text)That snippet follows the shape of OpenAI’s Responses API; confirm parameter names against the current docs, since the GPT-6 line is still moving. Three habits help. Start at low or medium effort and raise it only when a result misses. Keep the stable part of a prompt at the front so caching works. And set a rule for escalating to Sol when Luna fails, because a retry costs more than the saving.
GPT-6 Luna for Writing and Content Creation
Where it fits
Luna suits the small, repetitive pieces of a content workflow: meta descriptions, product blurbs, summaries of source material, outlines, tone adjustments and sorting posts into categories. OpenAI says its newer models give slightly shorter answers with fewer low-value details, as noted by Puter, which helps when you want tight copy and not padding. At about a tenth of a cent per short draft, you can afford to generate several options and pick one.
Where the evidence is thin
Nobody has published a credible writing-quality test of Luna that we could find. OpenAI’s guidance names structured summaries, not open-ended drafting. Artificial Analysis found weaker presentation and missed brief requirements in its knowledge-work tests. And the hallucination numbers above mean a confident paragraph can still be wrong. For a flagship article, use a stronger model or a human writer for the final pass, and check every fact, figure and date. Our 10-step AI content creation and fact-checking framework gives a reusable checklist for that.
A sensible split: let Luna handle research summaries, outlines and first-pass variants, then let a person (or Sol) own voice, claims and the final edit.
GPT-6 Luna Prompts to Try
OpenAI’s GPT-6 guidance says its sample prompts reflect behavior seen with Astra and should be evaluated on your own model and workload (OpenAI model guide). We haven’t tested the templates below; they are starting points built around what Luna is meant for, with a guard against guessing.
Structured summary
Summarize the text below for a reader with two minutes. Return a one-sentence takeaway, five bullet points, and every number or date exactly as written. If something is not in the text, write "not stated" instead of guessing.
Text: [paste text]Constrained rewrite
Rewrite the paragraph below at an eighth-grade reading level. Keep every fact, add no new claims, and stay under 90 words. Return only the rewrite.
Paragraph: [paste paragraph]Outline from notes
Using only the notes below, propose an outline with H2 and H3 headings for a 1,500-word article. Mark any section where the notes do not contain enough evidence.
Notes: [paste notes]Each prompt gives the model a narrow job, a clear output shape and permission to say “not stated”. That plays to Luna’s strengths and limits the damage when it doesn’t know something.
GPT-6 Luna Benefits and Limitations
| Benefits | Limitations |
|---|---|
| Very low price: $0.10 input, $0.50 output per million tokens | Independent scores are flat against GPT-5.6 Luna; the gain is price |
| Fast output, about 134 tokens per second at high effort | At max effort it can take well over a minute to start answering |
| Huge 1.05-million-token context window | Requests above 272,000 input tokens cost double on input |
| Free route through ChatGPT Free and Go | Access varies by plan and surface, and OpenAI’s own pages disagreed in October |
| Supports tools, structured outputs and caching | Uses more tokens per task than its predecessor, which narrows the savings |
| Strong cost per task on OpenAI’s coding and agent tests | Weaker than Sol on harder coding and knowledge-work tests |
| Cheap enough to generate and compare several drafts | Hallucination remains high on knowledge tests, so facts need checking |
Is GPT-6 Luna Worth Using? A Review of the Evidence
For high-volume, well-defined work, yes. Classification, extraction, summaries, routing and short drafts are what Luna is priced and tuned for, and at these rates you can run them at a scale that used to be uneconomical. For hard reasoning, serious coding or a finished long-form article, the evidence says step up to Sol or Astra, or bring in a human editor.
| If you are… | Verdict | Why |
|---|---|---|
| Running a high-volume pipeline | Good fit | Lowest price in the GPT-6 family, with caching and batch discounts |
| A free ChatGPT user | Good fit | It is what OpenAI gives Free and Go plans, with no extra cost |
| A blogger or content team | Partial fit | Useful for outlines, summaries and variants, not for unchecked final copy |
| Building agents | Good as a worker | Handles small steps while a larger model plans and checks |
| Doing complex coding or deep reasoning | Look higher | Sol and Astra score better on harder tests |
If you are also weighing another vendor’s latest, our Gemini 4 Argon breakdown covers Google’s side. Then re-check in a month: this lineup has already changed faster than most.
GPT-6 Luna FAQs
What is GPT-6 Luna?
GPT-6 Luna is OpenAI’s fastest and most cost-efficient GPT-6 model, released on September 22, 2026. It takes text and images, returns text, and is aimed at high-volume, well-defined tasks.
How much does GPT-6 Luna cost?
The API price is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.01. Batch jobs are half price, and requests above 272,000 input tokens cost more.
Is GPT-6 Luna free?
In ChatGPT, yes: Free and Go users can use it in the desktop app and, as it rolls out, in the Chat tab. API use is billed per token.
Is GPT-6 Luna better than ChatGPT?
The question mixes a product with a model. ChatGPT is the app, and Luna is one of the models inside it. A fairer comparison is Luna against GPT-5.6 Luna, Sol or Astra.
What is the difference between GPT-6 Luna and GPT-6 Sol?
Sol is built for complex coding and professional work and costs $2 and $10 per million tokens. Luna costs $0.10 and $0.50, runs faster, and scores lower on harder tests.
Can GPT-6 Luna write articles?
It can draft short pieces, outlines and summaries well enough to be useful. We found no strong evidence for long-form quality, and its hallucination figures mean every claim needs checking.
What is the GPT-6 Luna context window?
It is 1,050,000 tokens, with up to 128,000 tokens of output. Requests that pass 272,000 input tokens are billed at higher rates.
References
- OpenAI: Introducing GPT-6 Sol and Luna (accessed via search excerpts)
- OpenAI: A model guide for the GPT-6 family (accessed via search excerpts)
- OpenAI API docs: Latest model guide (accessed via search excerpts)
- Lindy: GPT-6 explained, updated October 1, 2026
- Puter: GPT-6 Sol and Luna review, September 25, 2026
- NeuralTrust: GPT-6 Sol vs Luna, October 6, 2026
- Unifically: GPT 6 Luna API model page
- Notebookcheck: ChatGPT Free and Go users start getting GPT-6 Luna in Chat, October 8, 2026
- Notebookcheck: ChatGPT Free gets GPT-6 Luna, but only in the desktop app, September 24, 2026
- OpenRouter: GPT-6 Luna model listing
- LLM Stats: GPT-6 Luna specs
- Artificial Analysis: GPT-6 Sol and Luna push the cost efficiency frontier (summary seen second-hand)
- Computing for Geeks: GPT-6 Sol and Luna hands-on (excerpt)
