- by x32x01 ||
OpenAI has expanded the GPT-6 family with two new models: GPT-6 Sol and GPT-6 Luna.
The biggest change is not only the new models themselves. OpenAI is also pushing the cost of advanced AI inference lower, making GPT-6 models more practical for developers building
According to OpenAI, GPT-6 Sol and Luna bring improvements from GPT-6 Astra into faster and more affordable models, while reducing their API prices by 50% compared with the promotional pricing of their GPT-5.6 counterparts.
Official source:
OpenAI - Introducing GPT-6 Sol and Luna
These are standard API prices per 1 million tokens. GPT-6 Sol also has a cached-input price of $0.20 per million tokens, while GPT-6 Luna is $0.01 per million cached input tokens.
For developers building an agent that needs to reason through multi-step tasks, Sol is positioned between the flagship GPT-6 Astra and the lower-cost Luna.
It is designed for workloads where the system may process a large number of requests and where API cost matters significantly.
Typical examples include:
OpenAI lists the following reductions compared with GPT-5.6 promotional pricing:
All prices are per 1 million tokens.
This matters because inference cost is part of the architecture of an AI application.
If an application calls a model thousands or millions of times, even a small reduction in the cost of each request can significantly affect the total operating cost.
💡 For developers, cheaper inference can mean more room for:
OpenAI says GPT-6 Astra remains its most capable model, while Sol and Luna are intended to bring more of the GPT-6 capabilities to applications operating under tighter cost or latency constraints.
In other words, developers do not necessarily need to use the most expensive model for every task.
A production system could use different models depending on how difficult each task is.
OpenAI also says it improved prompt caching for GPT-6, including higher cache hit rates and a 90% discount on cached input-token reads.
This can be especially useful for AI agents that repeatedly work with the same context.
For example, an agent working on a large software project may repeatedly send similar instructions, project information, or tool context. Reusing that context through caching can reduce both cost and latency.
OpenAI also provides controls for adjusting reasoning effort and tool availability while preserving earlier context for cache reuse.
Instead of sending every request to the most powerful model, applications can route tasks based on their requirements.
For example:
Simple task → GPT-6 Luna
More complex coding or agent task → GPT-6 Sol
Most demanding reasoning task → GPT-6 Astra
This type of model routing can help balance
⚙️ For large AI systems, that balance can be just as important as the model's raw benchmark performance.
They are also available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna through the desktop app. OpenAI's announcement also notes that the models were not yet available in regular ChatGPT conversations at the time of the announcement.
The more important change is the combination of:
The question is no longer only:
Which model is the most capable?
It is also:
Which model provides the right level of capability for the lowest practical cost per task?
That question becomes especially important as AI agents move from small experiments to systems that perform thousands or millions of tasks.
The biggest change is not only the new models themselves. OpenAI is also pushing the cost of advanced AI inference lower, making GPT-6 models more practical for developers building
AI agents, coding tools, automation systems, and high-volume applications.According to OpenAI, GPT-6 Sol and Luna bring improvements from GPT-6 Astra into faster and more affordable models, while reducing their API prices by 50% compared with the promotional pricing of their GPT-5.6 counterparts.
Official source:
OpenAI - Introducing GPT-6 Sol and Luna
GPT-6 Sol vs. GPT-6 Luna
The two models target different workloads:| Model | Best for | Input | Output |
|---|---|---|---|
| GPT-6 Sol | Complex coding and agentic workflows | $2 / 1M tokens | $10 / 1M tokens |
| GPT-6 Luna | Fast, high-volume, cost-sensitive tasks | $0.10 / 1M tokens | $0.50 / 1M tokens |
GPT-6 Sol
GPT-6 Sol is designed for more demanding work, especially:- Complex software engineering
- Coding agents
- Agentic workflows
- Professional tasks
- Computer-use workflows
- Applications that need a balance between capability and cost
For developers building an agent that needs to reason through multi-step tasks, Sol is positioned between the flagship GPT-6 Astra and the lower-cost Luna.
GPT-6 Luna
GPT-6 Luna focuses more heavily on efficiency and scale.It is designed for workloads where the system may process a large number of requests and where API cost matters significantly.
Typical examples include:
- High-volume text processing
- Classification
- Summarization
- Routine coding tasks
- Content transformation
- Automated workflows
- Large-scale AI agents
Why the API Price Cut Matters
The most interesting part of the release is the change in inference economics.OpenAI lists the following reductions compared with GPT-5.6 promotional pricing:
| Model | Previous Input | New Input | Previous Output | New Output |
|---|---|---|---|---|
| GPT-5.6 Sol → GPT-6 Sol | $4 | $2 | $20 | $10 |
| GPT-5.6 Luna → GPT-6 Luna | $0.20 | $0.10 | $1.20 | $0.50 |
This matters because inference cost is part of the architecture of an AI application.
If an application calls a model thousands or millions of times, even a small reduction in the cost of each request can significantly affect the total operating cost.
💡 For developers, cheaper inference can mean more room for:
- More agent iterations
- Larger workloads
- More automated tasks
- Higher usage limits
- More experimentation
- AI features that were previously too expensive to run at scale
GPT-6 Sol, Luna, and Astra
The GPT-6 family now covers different levels of capability and cost.| Model | Main focus |
|---|---|
| GPT-6 Astra | Highest-end reasoning and complex professional work |
| GPT-6 Sol | Complex coding and agentic workflows with lower cost |
| GPT-6 Luna | Efficient, high-volume workloads |
In other words, developers do not necessarily need to use the most expensive model for every task.
A production system could use different models depending on how difficult each task is.
Better Caching for AI Agents
The pricing changes are not the only efficiency improvement.OpenAI also says it improved prompt caching for GPT-6, including higher cache hit rates and a 90% discount on cached input-token reads.
This can be especially useful for AI agents that repeatedly work with the same context.
For example, an agent working on a large software project may repeatedly send similar instructions, project information, or tool context. Reusing that context through caching can reduce both cost and latency.
OpenAI also provides controls for adjusting reasoning effort and tool availability while preserving earlier context for cache reuse.
What This Means for AI Developers
The release points to a broader shift in how developers may design AI applications.Instead of sending every request to the most powerful model, applications can route tasks based on their requirements.
For example:
Simple task → GPT-6 Luna
More complex coding or agent task → GPT-6 Sol
Most demanding reasoning task → GPT-6 Astra
This type of model routing can help balance
quality, latency, and cost.⚙️ For large AI systems, that balance can be just as important as the model's raw benchmark performance.
GPT-6 Sol and Luna Availability
OpenAI says GPT-6 Sol and GPT-6 Luna are available through the API as: Code:
gpt-6-sol
gpt-6-lunaThe Bigger Change: AI Is Getting Cheaper to Run
GPT-6 Sol and Luna are not simply additional model names in the GPT family.The more important change is the combination of:
- Lower API prices
- Improved inference efficiency
- Better prompt caching
- Strong coding and agent capabilities
- More options for choosing a model based on workload
The question is no longer only:
Which model is the most capable?
It is also:
Which model provides the right level of capability for the lowest practical cost per task?
That question becomes especially important as AI agents move from small experiments to systems that perform thousands or millions of tasks.