GPT-6 Luna vs. GPT-5.6 Luna: Is the cheaper model good enough?

Written by
Last updated on:
November 16, 2026

GPT-6 Luna costs less than half as much as GPT-5.6 Luna, with comparable scores. It's slower, and neither model reaches our capability threshold, so the real question is which tasks they can handle.

Luna is OpenAI's low-cost model tier, built for high-volume work with a clear goal, like summarizing documents or answering short questions. OpenAI released GPT-6 Luna alongside GPT-6 Sol on September 22, 2026, with a price cut: $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for GPT-5.6 Luna.

That's less than half the price for comparable scores. But both Luna models sit below the capability threshold in our Ready for Enterprise (RFE) LLM Benchmark, so the sticker price isn't the only question. The real one is whether a budget model is good enough for the tasks you'd give it.

Should you switch from GPT-5.6 Luna to GPT-6 Luna?

For most high-volume tasks, yes. GPT-6 Luna costs less than half as much on a blended basis ($0.20 vs. $0.45 per million tokens), and its capability score is slightly higher (32.7 vs. 30.2).

The trade-offs are speed and a few benchmark results. GPT-5.6 Luna generates output about 30% faster in our measurements, and it scores slightly higher on four of the five benchmarks both models were tested on. If your workload is latency-sensitive, test both before you switch.

For harder steps, like multistep agentic work or complex coding, neither model is a strong fit. Route those to a more capable model and keep Luna for the simple, high-volume parts.

GPT-6 Luna vs. GPT-5.6 Luna: Key metrics

GPT-6 Luna vs GPT-5.6 Luna: RFE benchmark scores, pricing, speed, and context window compared
MetricGPT-6 LunaGPT-5.6 Luna
RFE rank#16#18
RFE score (out of 100)62.4 (below capability threshold)55.8 (below capability threshold)
Capability score (out of 100)32.730.2
Tested benchmark components6 of 65 of 6
Terminal-Bench 4.016.4%17.3%
SimpleQA Verified41.4%41.0%
Arena Text, style control14411453
Arena WebDev, Elo rating1579Not tested yet
GPQA Diamond90.5%91.6%
FrontierMath Tiers 1–378.9%82.1%
Price per million tokens$0.10 input / $0.50 output$0.20 input / $1.20 output
FullStack blended price per million tokens$0.20$0.45
Output speed70 tokens/second91 tokens/second
Context window1.05 million tokens1.05 million tokens

Source: FullStack RFE LLM Benchmark, US-hosted providers. Blended price assumes three input tokens for every output token. Neither model is on the SWE-Bench Pro leaderboard yet. Context windows as listed in OpenAI's model documentation.

Compare these models with others in our RFE benchmark.

What does "below the capability threshold" mean?

Our RFE score combines six public benchmarks with price and speed. A cheap, fast model can earn a high RFE score even if its benchmark results are modest. To keep that from pushing weak models to the top, we set a minimum capability score. Any model below it ranks after every model that reaches it. The capability score isn't a measure of general intelligence. It reflects how a model performs on those six benchmarks.

That's why GPT-6 Luna ranks 16th with an RFE score of 62.4, while models with lower RFE scores rank above it. Its price and speed earn a high score, but its capability score of 32.7 doesn't clear the bar.

In plain terms, being below the threshold doesn't mean a model is bad. It means it isn't a good default for complex work. For simple, well-defined tasks at high volume, a budget model can be the right choice, and that's what both Luna models are built for.

How do GPT-6 Luna and GPT-5.6 Luna compare on benchmarks?

In our benchmark, the results are close, and they point in different directions.

GPT-6 Luna has the higher capability score, 32.7 vs. 30.2. It's also the only one of the two tested on all six benchmarks. GPT-5.6 Luna hasn't been tested on Arena WebDev yet, so its score is based on five of the six.

On the five benchmarks both models were tested on, GPT-5.6 Luna scores slightly higher on four:

  • Terminal-Bench 4.0, which tests agentic tasks in a command-line environment: 17.3% vs. 16.4%.
  • Arena Text, a human-preference rating of written responses: 1453 vs. 1441.
  • GPQA Diamond, graduate-level science questions: 91.6% vs. 90.5%.
  • FrontierMath Tiers 1–3, advanced math: 82.1% vs. 78.9%.

GPT-6 Luna is slightly ahead on SimpleQA Verified, a factual accuracy test, at 41.4% vs. 41.0%. Most of these gaps are small enough that your own tests matter more than the benchmark differences.

Both models score low on Terminal-Bench 4.0, at under 18%. That's the clearest sign that neither is built for multistep agentic work.

What OpenAI reports about GPT-6 Luna

OpenAI's own launch results focus on how far Luna improved at higher reasoning effort. It’s worth flagging that these are OpenAI’s numbers, not independent tests, and many of OpenAI’s results were run at xhigh or max effort.

  • Business workflows. On AutomationBench, OpenAI reports that GPT-6 Luna at high effort scored 5.4 percentage points higher than GPT-5.6 Luna, at 58% lower cost per task.
  • Computer use. On the offline set of OSWorld 2.0, OpenAI says Luna at max effort beats GPT-5.6 Sol at medium effort, at about one-tenth the cost.
  • Honesty about its work. OpenAI's alignment tests found Luna made fewer misleading claims about its coding work than GPT-5.6 Luna.

OpenAI also notes that GPT-6 Sol and Luna give slightly shorter answers than earlier models. Treat that as a reason to measure, not a guarantee. A model can write shorter final answers and still spend more tokens on reasoning.

How much cheaper is GPT-6 Luna?

GPT-6 Luna's price cut is simple on paper: half the input price and less than half the output price. On a blended basis, it's $0.20 per million tokens compared with $0.45, or about 56% less.

At this tier, savings show up at volume. For example, take one task with 2,000 input tokens and 300 output tokens, like a short document summary:

  • GPT-6 Luna: about $0.00035 per request, or about $350 per million requests.
  • GPT-5.6 Luna: about $0.00076 per request, or about $760 per million requests.

Cached input tokens lower the bill further when requests reuse the same instructions or documents. OpenAI bills cached input reads at a 90% discount on GPT-6 Luna, and OpenAI notes that changing reasoning effort doesn't break the cache.

Why token prices aren't the whole task cost

Luna pricing per token is only part of the bill. Token usage per task matters just as much:

  • Reasoning effort. Reasoning tokens are billed as output tokens. GPT-6 Luna supports effort levels from none up to xhigh and max, with medium as the default. Higher effort can use many more tokens per task, which can erase the price cut.
  • Long prompts. Requests above 272,000 input tokens are billed at higher long-context rates.
  • Retries. A model with the lowest cost per token isn't always the cheapest per completed task if it needs more retries.

The practical metric is cost per accepted result on your own work. Run the same tasks on both models at the same effort level, and compare total token use as well as token prices.

For more ways to keep AI costs down, download our CIO's Guide to Controlling AI Coding Costs.

How does Luna pricing compare with other models?

For comparison, here's how list prices compare per million tokens with other OpenAI and Claude models:

GPT-6 Luna pricing compared with Claude Haiku 5.5, GPT-5.6 Luna, GPT-6 Sol, and Claude Sonnet 5.5: input, output, and blended price per million tokens
ModelInput priceOutput priceBlended price
GPT-6 Luna$0.10$0.50$0.20
Claude Haiku 5.5$0.10$0.50$0.20
GPT-5.6 Luna$0.20$1.20$0.45
GPT-6 Sol$2.00$10.00$4.00
Claude Sonnet 5.5$2.00$10.00$4.00

Prices from OpenAI and Anthropic published pricing as of November 2026. Claude Haiku 5.5 prices apply to requests under 100,000 tokens. Blended price assumes three input tokens for every output token.

GPT-6 Luna shares the lowest cost in this table with Claude Haiku 5.5, so for teams open to Claude models, the choice depends on how each model's performance holds up on your tasks. Premium models like GPT-6 Sol cost about 20 times as much on a blended basis, which is why routing simple steps to a budget model saves so much.

What does the slower speed cost you?

GPT-5.6 Luna is the faster model, at 91 tokens per second compared with 70 for GPT-6 Luna in our measurements. Here's what that means for estimated response time:

GPT-6 Luna vs GPT-5.6 Luna: estimated response times for short, detailed, and long responses
Response lengthGPT-6 Luna (estimated)GPT-5.6 Luna (estimated)
200 tokens (short answer)About 5 secondsAbout 4 seconds
1,000 tokens (detailed answer)About 17 secondsAbout 13 seconds
4,000 tokens (long report)About 60 secondsAbout 46 seconds

Estimates use time to first token plus output length divided by measured output speed. Real response times also depend on reasoning effort, tool calls, and load.

For batch jobs, like classifying a backlog of tickets overnight, the difference rarely matters. For real-time features, like a chat assistant that streams long answers, users may notice it. Test GPT-6 Luna on your real response lengths before you move user-facing features.

Where can you use GPT-6 Luna?

Through the OpenAI API, GPT-6 Luna is available as gpt-6-luna, with the same 1.05-million-token context window as GPT-5.6 Luna. It accepts text and image input.

In ChatGPT, OpenAI made it available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users can use it in the desktop app. Those subscription models are a separate product from the API, so usage in ChatGPT doesn't affect your API bill.

OpenAI doesn't publish GPT-6 Luna's model weights, so you can't run it alongside local models on your own hardware. You use it through a hosted API or OpenAI's apps.

Is GPT-6 Luna a drop-in replacement?

Close, but test before you switch. Both models use the same OpenAI API, so the change can be as small as a new model name. The behavior isn't identical, though. Answers may be shorter, results can shift at each effort level, and tool calls may behave differently in your workflows. Retest the steps most relevant to your business:

  • Build a small evaluation set from real requests, including the edge cases that cause problems today.
  • Run both models on the same prompts with the same reasoning effort.
  • Compare accuracy, latency, and cost per accepted result, not only token prices.
  • Keep the model name in configuration so you can switch back quickly if quality drops.

When is a budget model good enough?

Budget models like Luna work best on high-volume tasks with a clear goal and an easy way to check the result. Good candidates include:

  • Classification and routing, like tagging support tickets or sorting incoming requests.
  • Extraction, like pulling fields from invoices or forms into a structured format.
  • Short summaries of documents, calls, or threads.
  • Short answers to well-scoped questions, especially with retrieval from your own data.

They're a weaker fit for multistep agentic work, or for anything where a wrong answer is expensive and hard to catch. Even with fewer misleading claims, keep people reviewing code that a budget model writes.

For most teams, the practical setup is a mix. Send simple, high-volume steps to a budget model, and route harder steps to a more capable model, like GPT-6 Sol or GPT-6.1 Sol, which OpenAI announced at Dev Day a week after GPT-6 Sol. That way, you pay premium prices only where you need them.

Which Luna model should you use?

For most high-volume tasks, GPT-6 Luna is the better choice. It costs less than half as much as GPT-5.6 Luna, and its scores are close enough that the savings will usually matter more than the differences. Keep GPT-5.6 Luna where response speed matters, and route agentic work and complex coding to a stronger model.

The bigger savings come from that routing. When a budget model handles the simple steps and a premium model handles only the hard ones, you pay premium prices for a fraction of your traffic. 

Our AI cost optimization team helps companies do this. We start by measuring what your current AI usage costs and testing cheaper models on a sample of your real work. If a cheaper model holds up, we help you set up routing and track costs by task, so you can see where the savings come from.

Learn more about our AI cost optimization services.

Learn more

Frequently asked questions

Yes. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for GPT-5.6 Luna. That's $0.20 vs. $0.45 per million tokens on a blended basis, or about 56% less.

It's close. GPT-6 Luna has a slightly higher capability score in our RFE benchmark (32.7 vs. 30.2), but GPT-5.6 Luna scores slightly higher on four of the five benchmarks both were tested on, and it's faster. For most simple, high-volume tasks, GPT-6 Luna's lower price makes it the better value.

It can, depending on the task and the reasoning effort you choose. Higher effort levels like xhigh and max use more tokens, and reasoning tokens are billed as output. Measure token usage on your own tasks before assuming the lower price means a lower bill.

Our RFE benchmark ranks any model below a minimum capability score after every model that reaches it, even if its price and speed earn a high RFE score. Both Luna models fall below it. They can still be a good fit for simple, well-defined tasks, but they aren't a good default for complex work.

Route multistep agentic work and complex coding, plus any task where errors are costly or hard to catch. Both Luna models score under 18% on Terminal-Bench 4.0, which tests agentic tasks, so a more capable model is a safer choice for that work.