OpenAI has expanded its GPT-6 family with GPT-6 Sol and GPT-6 Luna, two models designed to bring more capable AI to coding, professional work and high-volume automation at lower API prices.
The two models sit below GPT-6 Astra, OpenAI’s flagship system. Sol is aimed at demanding coding and agentic workflows, while Luna is positioned for focused, repeatable tasks where cost and scale matter more. OpenAI released both models on September 22, 2026.
For developers, the most significant change may not be the model names or benchmark scores. It is the lower cost of repeatedly running AI inside real applications.
How GPT-6 Sol and Luna Differ
OpenAI’s current model guidance gives each GPT-6 model a different role.
GPT-6 Astra is intended for the hardest end-to-end reasoning and coding tasks. GPT-6 Sol is designed to balance intelligence and cost, particularly for complex coding and agentic workflows. GPT-6 Luna is the most cost-efficient option for focused, high-volume work.
That makes the choice less about finding one model for everything and more about matching the model to the workload.
A simple automated process may not need the most expensive model. A complicated coding task involving multiple tools and iterations may justify Sol or Astra.
GPT-6 Sol and Luna Pricing
The standard API prices are substantially different between the two models:
| Model | Input / 1M tokens | Cached input / 1M | Output / 1M tokens |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
These are OpenAI’s standard processing prices for shorter contexts. Longer-context requests have higher rates, so developers should not assume that every API request will use the figures above.
OpenAI says the new prices represent a 50% reduction compared with the promotional pricing previously offered for GPT-5.6 Sol and Luna. That is an OpenAI pricing comparison, not a guarantee that every application will cost exactly 50% less overall.
The distinction matters because an application’s final bill depends on how much input and output it generates, how often it calls the model and how effectively it uses caching.
Why the Lower Cost Matters for Coding Agents
AI coding agents often perform several operations before completing a task. They may inspect files, reason about changes, call tools, modify code and test the result.
That repeated interaction can make token costs accumulate quickly.
OpenAI reports that GPT-6 Sol reaches 68.8% on DeepSWE v1.1, while GPT-6 Luna reaches 66.6% at maximum reasoning effort. These are OpenAI-reported evaluation results, so they should be treated as company benchmark claims rather than independent evidence of real-world performance.
OpenAI also reports a 60.5% result for GPT-6 Sol at xhigh effort on the offline OSWorld 2.0 evaluation, compared with 60.3% for Claude Opus 5 at medium effort, with OpenAI reporting a substantially lower cost per task for Sol. Again, these comparisons depend on the specific evaluation setup and should not be interpreted as universal performance rankings.
For developers, the more useful takeaway is that lower-cost models can make repeated AI-assisted operations easier to run within a fixed budget.
Caching Can Reduce Repeated Processing
Model pricing is only part of the economics.
OpenAI says GPT-6 Sol and Luna preserve earlier context for cache reuse even when reasoning controls change. Developers can also use explicit cache breakpoints to control which prompt prefixes remain reusable.
Caching is useful when an application repeatedly sends the same large instructions or context. Instead of processing the entire repeated prefix as new input every time, eligible cached input can be billed at a lower rate.
OpenAI says improvements to its caching system have reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to its models. That is an OpenAI-reported operational figure, not an independent industry measurement.
For developers building long-running agents, this can matter because a small saving on every repeated request can become significant at large volumes.
Who Should Use Sol or Luna?
The models are available through the OpenAI API, with model IDs gpt-6-sol and gpt-6-luna. They are also available in ChatGPT Work and Codex for eligible paid plans. OpenAI says Free and Go users can access Luna through the desktop app, while the models were initially separate from the standard Chat experience.
For a developer, the practical choice depends on the workload.
Luna makes more sense when an application needs large numbers of relatively focused operations and cost efficiency is important.
Sol is better suited to more demanding coding and agentic tasks where stronger reasoning can justify the additional cost.
Astra remains the option OpenAI recommends for the hardest end-to-end work.
What Readers Should Know
GPT-6 Sol and Luna are best understood as different cost-and-capability options within the same model family, rather than simply two cheaper versions of Astra.
Sol targets demanding coding and agentic workflows, while Luna focuses on efficient, high-volume work. Their API prices are substantially lower than Astra’s, and OpenAI has also improved caching and inference efficiency.
The important limitation is that benchmark results and performance claims come largely from OpenAI’s own evaluations. Real-world results can vary depending on prompts, tools, workload and application design.
For developers, the key question is therefore not simply which model is most powerful. It is which model can complete the required task reliably at a reasonable total cost.
