GPT-6 Luna: what OpenAI's cheapest new model is, and who it's for

What is GPT-6 Luna?

GPT-6 Luna is the smallest of the three models in OpenAI's GPT-6 lineup, which OpenAI rolled out in stages: the GPT-6 Astra flagship on 3 September 2026, then GPT-6 Sol and GPT-6 Luna together on 23 September. Luna sits at the bottom of that family by price and by weight, the way GPT-5.6 Luna did in the last generation.

Think of it as the model you point at work you need to do a lot of, cheaply and fast, where the task is not especially hard. Classification, extraction, routing, summarising, simple drafting, the high-volume plumbing of an AI product. It is not the model you reach for to reason through a thorny problem or drive a long, multi-step agent. That is what Sol and Astra are for.

The specs back that positioning up. GPT-6 Luna accepts text and image input and returns text, with a large 1,050,000-token context window, up to 922,000 tokens of input and 128,000 tokens of output per request. Its knowledge cutoff is 18 May 2026, a month later than Sol's. It exposes adjustable reasoning effort from none all the way to max, with medium as the default, and it ships the full Responses API tool set: web search, code interpreter, hosted shell, computer use, and MCP. In other words, it is a fully featured modern model, just tuned for cost and throughput rather than raw horsepower.

What GPT-6 Luna is good at

The cleanest way to read Luna is OpenAI's own tagline: "Build with Sol. Scale with Luna." You develop and handle your hard cases on the mid-tier model, then run your high-volume, latency-tolerant, simpler traffic on the cheap one. Luna is the second half of that sentence.

That framing holds up against the benchmarks, which are worth being honest about because a big "6" in the name suggests a capability jump that mostly is not there. Artificial Analysis reported that GPT-6 Sol and Luna "push the cost efficiency frontier by halving cost," while the Intelligence Index and Coding Agent Index "remain level with GPT-5.6, with progress in some evaluations and regressions in others." So the intelligence is steady and the price dropped.

Where Luna did improve is reliability. OpenAI says the GPT-6 family makes fewer factual mistakes than its GPT-5.6 predecessors, and Artificial Analysis logged Luna's max-effort hallucination rate dropping from 93% to 77%. That is a real improvement for a light model, though still higher than Sol's on the same test, which is exactly why you would not put Luna on your hardest, highest-stakes questions.

The community read it the same way. On the launch discussion, one developer summed the trade up:

"I'd have loved improved performance for 2x the price than ~equal performance for 0.5x the price."

dmazin, Hacker News

So the honest shape of Luna: a dependable, cheap, fast light model. Perfect for volume, wrong for your hardest reasoning.

What GPT-6 Luna costs

The price is the headline, so here is the short version, with the full tier-by-tier detail in the GPT-6 Luna pricing guide.

On the API pricing page, GPT-6 Luna standard pricing is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.01. That makes it the cheapest model in the family by a wide margin.

Source: OpenAI API pricing. Two things stand out. First, Luna is one fifth the price of GPT-6 Sol and one hundredth of GPT-6 Astra on output. Second, it is a straight cut from the prior generation: GPT-6 Luna is the old GPT-5.6 Luna with input halved and output cut by well over half. OpenRouter confirmed the rates on launch day.

One thing that quietly reshapes a budget: Luna runs on the same service tiers as the rest of the family. Batch and Flex run at $0.05/$0.25, half of standard, while Fast mode doubles it to $0.20/$1.00. For a high-volume model, that service-tier choice is the single biggest lever on your bill.

How to access GPT-6 Luna

This is where Luna's story is different from the rest of the family: it is the one everyone gets. There are two routes.

In ChatGPT and Codex. GPT-6 Luna is included on every plan, from Free up. Here is how the plans map:

Source: OpenAI launch announcement. The headline for non-developers is that the Free plan and the $8 Go plan get GPT-6 Luna in the desktop app, and neither gets Sol. Paying more does not take Luna away; it adds Sol on top. OpenAI shipped both models into Codex at the same time, so coding-agent users can pick Luna for cheaper, lighter runs too.

On the API. Developers call the gpt-6-luna model ID through the OpenAI API. The rate-limit floor is Tier 1 at 500 requests and 500K tokens per minute, scaling to 30,000 requests and 180M tokens per minute at Tier 5. Everything is billed per token, at the rates above.

One caveat worth flagging from the launch thread: a few developers noted that message-limit changes suggested the 50% API price cut translated to something closer to a one-third saving inside the ChatGPT subscriptions, so check your own plan's allowance rather than assuming the sticker discount flows straight through.

The catch: where GPT-6 Luna stops

Luna is a good deal, but a few things can trip you up if you do not know them going in.

It is not for hard reasoning. This is by design. Luna posted only small gains on the family's agentic evals, and its hallucination rate, while improved, is the highest of the three GPT-6 models. If your task needs careful multi-step reasoning or long-horizon tool use, Luna is the wrong tier, and the answer is Sol or Astra.

Long context is billed at a higher rate for the whole request. Luna's window is a large 1,050,000 tokens, but any prompt over 272K input tokens flips the entire request to the long-context rate: $0.20 per million input and $0.75 per million output, applied to every token, not just the ones past the threshold. It is easy to miss when you budget for large-context work.

Cheap tokens are not the whole cost of shipping something. A rock-bottom token price is a discount on the cheapest part of any real AI project. The expensive part is everything around the model: connecting it to your data, testing it, making it safe, and keeping it accountable. More on that below, because for one big use case it changes the whole calculation.

Which GPT-6 model should you use?

If you are choosing between the three, the decision is mostly about how hard your task is and how much of it you run. Pick your main job and see where it lands.

For most people the honest answer is that you will use Luna in ChatGPT without thinking about it, and reach for Sol on the API when you are building something that needs to be dependable under pressure.

What GPT-6 Luna means if you run a support team

Here is where I want to be useful rather than just recite specs. I build AI agents at eesel, so most of my time goes to the gap between a model and a working system, and for support teams that gap is the whole ballgame.

If you run customer support or internal ops, a cheaper frontier model changes less than the headline suggests. A raw model like Luna is infrastructure, not an employee. It is a capable, cheap engine, but out of the box it does not know your refund policy, it cannot see your last 50,000 tickets, it has no access to your order system, and it has no safe way to be tested against your real queue before it starts replying to customers.

The unit that actually matters for a support team is cost per resolution, not cost per token. A token bill is unpredictable: it moves with prompt length, retries, and how chatty the model is on any given day. Across years of putting AI agents on live support queues, I have watched confident-sounding bots quietly give wrong answers, which is exactly why every rollout should be simulated against your own historical tickets before it goes anywhere near a customer. That gap, between a model and a system built around it, is your knowledge, your integrations, safe actions, and a way to test on your history first.

It is also why I would not re-platform your support automation the week a new model ships. The best AI helpdesk agents and customer service tools are model-flexible by design, so when a cheaper or smarter model lands, you inherit the upgrade without rebuilding anything.

Try eesel

If your real question about Luna is "what will an AI cost me to actually resolve tickets," that is the job an AI teammate is built for, and it is where eesel sits. GPT-6 Luna is the engine; eesel is the employee you hire to drive it. The AI helpdesk agent trains on your past tickets and help center, plugs into the helpdesk you already run instead of replacing it, and lets you simulate it against thousands of your real historical tickets so you see how it will perform, and what it will cost, before it touches a live conversation.

The billing tells the story a token rate cannot. Instead of paying per token to think, you pay 40¢ per ticket it actually handles, with no per-seat fees and no platform minimum, so the cost tracks outcomes rather than model chatter. You can start free with $50 of usage and no credit card.

And because this is often a developer's question, it is worth knowing the whole eesel workspace is drivable from the terminal, not just a dashboard. The eesel CLI (npx @eesel/cli) lets a person, a script, or a coding agent like Claude Code, Codex, or Cursor operate the same teammate: connect a helpdesk, read and edit the standing instructions, run eesel activity to inspect past runs, and gate changes behind eesel approvals. Every command prints JSON, and --dry-run shows the exact server call a write would make before it sends. So the per-resolution model is something you can wire into your own tooling, the same way you would call gpt-6-luna from a script.

GPT-6 Luna is the cheapest way to run a capable model right now, and for high-volume, simple work the price is worth taking. For most support teams, though, the token rate was never the number that mattered. Getting a tested, connected, accountable teammate into your existing workflow, billed by the outcome, is the part worth spending your budget on.