GPT-6 Sol and GPT-6 Luna arrived in the OpenAI API on September 22, 2026 with a large price gap and a relatively clear usage split. Under Standard processing, for requests with up to 272K input tokens, GPT-6 Sol costs $2 per 1 million input tokens and $10 per 1 million output tokens. GPT-6 Luna costs $0.10 for input and $0.50 for output for the same volume.
That makes Luna 20 times cheaper per token than Sol at this pricing tier, for both input and output. The cost difference is objective, but it does not make Luna the automatic choice. OpenAI positions Sol for complex coding and agentic workflows, while it describes Luna as its most efficient model for focused, high-volume tasks. The practical decision depends on how much any additional quality from Sol changes the outcome of your workflow.
GPT-6 Sol and Luna API prices
| Model | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| GPT-6 Sol | $2.00 | $0.20 | $2.50 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 |
The figures above are per 1 million tokens, under Standard processing and short context, currently defined in the documentation as up to 272K input tokens. OpenAI's pricing page and model cards also list different rates for long context and other processing modes. See the official OpenAI pricing table, plus the model pages for GPT-6 Sol and GPT-6 Luna.
How much GPT-6 cut prices versus GPT-5.6
For GPT-5.6 Sol, the promotional reference used by the launch announcement was $4 per million input tokens and $20 per million output tokens. GPT-6 Sol halves both figures to $2 and $10.
For GPT-5.6 Luna, the reference was $0.20 per million input tokens and $1.20 per million output tokens. GPT-6 Luna lowers input to $0.10 and output to $0.50. Input is therefore 50% cheaper and output is about 58% cheaper. The launch announcement summarizes the new generation as 50% cheaper than the previous promotional pricing, while the detailed table provides the exact figures by category.
A hypothetical workload cost example
Consider an application that sends 10 million input tokens and receives 2 million output tokens in one month, with no cache usage and no request exceeding 272K input tokens.
- GPT-6 Sol: 10 x $2 + 2 x $10 = $40.
- GPT-6 Luna: 10 x $0.10 + 2 x $0.50 = $2.
In this example, Luna reduces the token bill from $40 to $2, a $38 difference. Because Luna's input and output rates are both 20 times lower than Sol's at this tier, the same cost ratio holds for any equivalent mix of input and output tokens as long as both workloads are subject to the same processing rules.
When GPT-6 Sol makes more sense
OpenAI describes Sol as a model built for complex coding and agentic workflows. That points to scenarios where mistakes are costly, a task has many steps, tool use is intensive, or the result needs to sustain more demanding reasoning and execution.
- agents that plan and execute long sequences of actions;
- complex coding, code review, and repository-wide changes;
- professional tasks where a small quality improvement can reduce human rework;
- workflows where you have already measured enough quality gain to justify the higher cost.
The reason to pay more should come from testing on your own task set. Without an internal benchmark, the official positioning should not be turned into a promise that Sol will be better for every application.
When GPT-6 Luna makes more sense
Luna is presented as the most efficient model for focused, high-volume tasks. Its price favors applications that make many calls, process large batches, or have low unit value per request.
- classification, extraction, and structured text transformation;
- summaries and short answers with well-defined criteria;
- pre-processing before harder cases are routed to a more expensive model;
- large-scale operations where cost and throughput are central to the decision.
One architecture worth testing is to use Luna by default and route only harder cases to Sol. This is not a universal recommendation: it needs a measurable routing rule and validation to prevent a silent quality drop.
The two models share several technical limits
The official model cards list a 1.05 million token context window and a 128K maximum output for both Sol and Luna. Both accept text and image input and generate text. They also support reasoning effort settings from none through max, although actual cost and latency depend on real usage and the selected processing mode.
These similar limits can make architectural migration easier, but they do not remove the need for quality testing. The same context window does not imply the same performance in reasoning, coding, tool use, or instruction following.
Long context, caching, and processing modes change the bill
The $2/$10 Sol rates and $0.10/$0.50 Luna rates do not cover every case. According to the documentation, prompts above 272K input tokens are charged, for the entire request, at 2 times the input and cache rates and 1.5 times the output rate.
Caching also has its own rates. Cached input costs $0.20 per million tokens for Sol and $0.01 for Luna. Cache writes cost $2.50 for Sol and $0.125 for Luna. OpenAI also states that Batch and Flex are priced at 50% of Standard rates, while Fast mode is priced at 2 times the applicable rate. Regional processing can add a 10% premium where available.
How to choose without your own benchmark
If you do not yet have an evaluation set, start by separating tasks by risk and value. For high-volume, low-impact calls, Luna has a very large cost advantage. For complex workflows, especially coding and agents, Sol is the model OpenAI positions for that type of work.
Then build an internal benchmark with real examples and measure at least four items: success rate, human correction required, latency, and total cost per completed task. The useful metric is not only price per million tokens, but the cost to reach an acceptable result.
What the price comparison does not prove
This comparison uses official prices and positioning descriptions published by OpenAI on September 22 and 23, 2026. It does not include an independent benchmark produced for this article. Prices can establish how much a given token volume costs under equivalent rules, but they cannot, by themselves, establish which model will deliver better quality, lower latency, or a higher success rate in your application.
For a safer migration, the choice between GPT-6 Sol and Luna should combine the cost table with tests on data that represents your product. Luna provides the lower token-cost route; Sol is aimed at more complex work. The cutoff between them should be defined by the results of your own system.
