GPT-6 Luna's tokens cost a twentieth of GPT-6.1 Sol's. On Artificial Analysis a finished task costs about half, and Sol on low effort scores five points higher than Luna on max.
Ten days in, GPT-6 Luna has enough independent results to price properly. The sticker says it is a twentieth of Sol. The work says something closer to half, and a model from Xiaomi says Luna isn't the floor.
Photo by Krzysztof Kowalik on Unsplash
Short version: Luna is the budget tier of GPT-6 and half the price of the Luna it replaces. Run at max effort it ties GPT-5.6 Luna for 61% less per task. GPT-6.1 Sol on low effort beats it by five points for 1.9x the money, and MiMo-V2.6-Flash beats it for less.
A twentieth of Sol, line by line
OpenAI shipped Luna on September 22 next to GPT-6 Sol. Input and output are exactly 1/20th of GPT-6.1 Sol's $2 and $10. Cached input is the one line where the ratio breaks: Sol 6.1 halved its cache read to $0.10 on September 29, so the gap there is 10x, not 20x.
| Per 1M tokens | Input | Cached | Output | Over 272K |
|---|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | $0.20 / $0.75 |
| GPT-6 Luna Batch | $0.05 | $0.005 | $0.25 | $0.10 / $0.375 |
| GPT-6 Luna Fast | $0.20 | $0.02 | $1.00 | $0.40 / $1.50 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | $0.40 / $1.80 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $4.00 / $15.00 |
OpenAI pricing page and GPT-6 Luna model page, read October 2, 2026. Over-272K column is input / output for the whole request. Cache writes are $0.125 standard. Flex matches Batch on short prompts; OpenAI lists no Flex long-context rate.
Against its own predecessor, Luna halves input and takes 58% off output. The detail we like most is the Fast row. Luna on OpenAI's priority tier costs $0.20 and $1.00, which is cheaper on output than GPT-5.6 Luna at standard speed. There's no Ultrafast tier for Luna; that launched on September 29 for GPT-6 Astra only.
Per task, 20x turns into 1.9x
Token prices only tell you the cost of a token. Reasoning models decide how many they spend, so we look at what a finished task costs. Artificial Analysis runs every model through the same index and publishes the bill.
- 48GPT-6.1 Sol medium$0.21
- 42GPT-6.1 Sol low$0.13
- 41Gemini 3.8 Flash high$1.24
- 39DeepSeek V4.1 Flash max$0.27
- 38MiMo-V2.6-Flash$0.06
- 37GPT-6 Luna max$0.07
- 37GPT-5.6 Luna max$0.18
- 29GPT-6 Luna medium (default)$0.02
- 17Claude Haiku 4.5 reasoning$0.28
Score on the Artificial Analysis Intelligence Index v4.3.2 and cost per index task, from the leaderboard, read October 2, 2026. Luna's own model page shows 38 at max.
Luna at max scores 37 for $0.07. Sol at low scores 42 for $0.13. That's 1.9x the cost per task for five more points, from a model whose tokens cost 20 times as much. Sol on low just doesn't think very long. Luna on max thinks a lot, and on the model page Artificial Analysis clocks it at 124 seconds to first token, against 2.6 seconds for Sol on low.
So if you were planning to run Luna at max to squeeze out quality, don't. Sol on low is better and faster for less than double. Luna's case is the default setting, medium: 29 for $0.02 a task, which is a different price class from everything else on this chart.
The effort dial inside Luna
Luna takes six effort settings: none, low, medium, high, xhigh and max. Medium is the default. Going from medium to max takes a task from two cents to seven and buys eight points. The middle steps sit roughly on a line between them: high is 32 for $0.03 and xhigh is 34 for $0.04.
With reasoning off it scores 18 for about a cent a task. That beats Claude Haiku 4.5 with reasoning on, which scores 17 for $0.28. We don't think many people are still choosing Haiku 4.5 on price, but if you are, you're paying 28 cents for what Luna does for about one.
We left low effort off the chart on purpose. The leaderboard lists it at 21 for $0.0045, cheaper than running with reasoning off, which doesn't make sense for the same model. We'd wait for that row to be rerun before trusting it.
Xiaomi got there first
MiMo-V2.6-Flash came out on September 21, a day before Luna, with open weights under MIT. It lists at $0.14 in and $0.28 out. On Artificial Analysis it scores 38 for $0.06 a task, one point over Luna for about 8% less. The Vals Index makes the gap wider.
| Model | Vals Index | Cost per test |
|---|---|---|
| GPT-6.1 Sol | 61.15% | $3.24 |
| Gemini 3.8 Flash | 54.83% | $5.73 |
| MiMo-V2.6-Flash | 53.23% | $0.20 |
| DeepSeek V4.1 Flash | 51.32% | $0.33 |
| GPT-6 Luna | 51.22% | $0.43 |
| GPT-5.6 Luna | 51.69% | $0.82 |
Vals AI index v2.1, updated September 30, 2026.
On Vals, MiMo scores two points higher than Luna for 53% less per test, and DeepSeek V4.1 Flash matches Luna for 23% less. Luna's win here is over its own predecessor: the same score as GPT-5.6 Luna for 48% less. Sol 6.1 charges $3.24 a test, $2.81 more than Luna, for about ten more points.
MiMo has a real catch, and it isn't the price. Artificial Analysis measures it at about 54 output tokens a second against Luna's 127, and Vals logs its latency at 58 minutes against Luna's 31. If a user is waiting on the answer, that matters more than a cent. If it's a batch job, it doesn't. We also couldn't load Xiaomi's own pricing page, so the $0.14 and $0.28 come from OpenRouter and Artificial Analysis.
Small print that changes the bill
The 272K line. A prompt of 272,000 tokens costs $0.0272 to send. A prompt of 273,000 costs $0.0546, because past the line the whole request bills at $0.20 input and $0.75 output, not just the extra thousand tokens. Luna's context goes to 1,050,000, so plenty of RAG and document jobs will land on the far side of it.
Bedrock sells Luna at OpenAI's list price on its global endpoint and 10% more on US regional ones: $0.11 and $0.55. OpenAI charges the same 10% for its own regional processing. Azure's pricing page still says Luna's prices are being processed.
And if you send images: on September 25 OpenAI fixed a bug in image encoding that had degraded image understanding in both GPT-6 Sol and Luna, and told users to rerun their evals. Any vision test you ran on Luna in its first three days measured a broken model.
Medium for volume, Sol for quality
Luna on medium or with reasoning off is the right default for high-volume, low-stakes work: classification, extraction, routing, short replies. At two cents a task or less there isn't much left to save, and the batch tier halves that again.
Once you find yourself turning Luna up to xhigh or max, stop and price Sol on low instead. On the numbers we have it scores higher, answers far sooner and costs less than twice as much per task. And if latency doesn't matter and open weights do, MiMo-V2.6-Flash is cheaper than both.
Put your own token mix into the calculator to see Luna, Sol and the Flash models side by side.
Pages we checked
- OpenAI: API pricing and GPT-6 Luna model page - every tier, long-context rates, context, effort settings
- OpenAI: API changelog - September 22 release, September 25 image-encoding fix, September 29 Ultrafast
- Artificial Analysis: leaderboard and GPT-6 Luna - Intelligence Index scores, cost per task, speed, time to first token
- Vals AI: Vals Index - accuracy, cost per test, run time
- AWS: GPT-6 Luna on Bedrock - global and regional prices
- DataNorth: Xiaomi releases MiMo-V2.6 - release date and licence