GPT-6 Sol and Luna are the two AI models OpenAI released on September 22, 2026, and they cut API prices in half. Sol targets demanding coding and reasoning work. Luna handles routine tasks like summarization and data extraction. Together, they replace GPT-5.6 Sol and Luna as the default models in ChatGPT, Codex, and the OpenAI API. For developers running high-volume AI workloads, the price cut changes real monthly bills, not just benchmark scores.
What Are GPT-6 Sol and Luna
GPT-6 Sol is OpenAI’s model for demanding professional work. It handles coding, feature development, code review, and debugging. GPT-6 Luna is the lower-cost option for high-volume tasks. It summarizes documents, extracts data, and answers straightforward questions.
Both models became available on September 22, 2026, replacing GPT-5.6 Sol and Luna. They ship in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts. Luna also rolled out to the ChatGPT desktop app for Free and Go users. In the API, developers can call them directly as gpt-6-sol and gpt-6-luna.
OpenAI has been building toward task-specific coding models for a while. Its software engineering agent reflected an early bet. OpenAI bet that specialized coding tools, not just bigger general models, would drive adoption. Sol extends that bet with a model priced for everyday use, not just demos.
Why OpenAI Cut Prices in Half
Sol costs $2 per million input tokens and $10 per million output tokens. That is half the price of GPT-5.6 Sol, which cost $4 and $20. Luna costs $0.10 per million input tokens and $0.50 per million output tokens. That’s roughly half of its predecessor’s pricing.
OpenAI credits the cut to improvements in inference and caching. Cached token reads now get a 90 percent discount. That matters most for agents that repeatedly reread the same context. Think of a code review bot scanning one repository all day. For those workloads, the caching discount often affects your bill more than the headline price. Check your own cache hit rate before you assume the sticker price tells the full story.
How GPT-6 Sol Performs on Coding Benchmarks
OpenAI published benchmark comparisons against Anthropic’s Claude models. On AutomationBench, Sol scores 33.2 percent at its highest effort setting. OpenAI says that beats Claude Opus 5 at roughly 9 percent of the cost.
On Agents’ Last Exam, Sol reaches 56.4 percent at maximum effort. That result also edges out Claude Opus 5, at 60 percent lower cost per completed task. On the coding-focused DeepSWE v1.1 benchmark, Sol scores 68.8 percent. That is within 1.1 points of Anthropic’s Claude Fable 5, at 80 percent lower cost. OpenAI picked Fable 5 as the comparison model for that specific test.
These are OpenAI’s own published numbers, not independently audited results. Treat them as a starting point for evaluation, not a final verdict. Full methodology and effort settings sit in OpenAI’s announcement post.
Choosing Between Sol and Luna for Your Team
A team building a code review bot that runs on every pull request is a Sol use case. The task needs real reasoning, and the cost per run stays low even at Sol’s higher per-token price.
A support tool that classifies incoming tickets and extracts key fields is a Luna use case. The task is repetitive and high-volume, and Luna’s lower price matters more than deep reasoning. Running that same classification job on Sol would work. It would just cost roughly 20 times more per token for little quality gain.
Mixed pipelines are common too. A documentation generator might use Luna to summarize source files, then hand those summaries to Sol for structured output. That split keeps the expensive model working only on the step that actually needs it.
Tradeoffs and Common Mistakes to Avoid
The biggest mistake is treating a 50 percent price cut as a 50 percent budget cut. Cheaper tokens often lead teams to run more agent loops, more retries, and more parallel tasks. Total spend can climb even as unit cost drops.
The second mistake is skipping model selection entirely and defaulting every task to Sol. Luna handles most production workloads at 20 times lower cost per token. Reserve Sol for tasks where its extra reasoning actually changes the output.
Vendor benchmarks deserve scrutiny before they inform architecture decisions. Run your own evaluation set against your actual prompts and data before switching models. Teams already updating their software development practices for AI should treat this release as one update worth testing. It is not a reason to rearchitect a pipeline overnight.
Key Takeaways
- GPT-6 Sol and Luna launched September 22, 2026, replacing GPT-5.6 Sol and Luna at half the price.
- Sol costs $2 and $10 per million input and output tokens; Luna costs $0.10 and $0.50.
- OpenAI’s own benchmarks show Sol matching or beating Claude Opus 5 at 9 to 60 percent of the cost, depending on the test.
- Luna suits high-volume, repetitive tasks; Sol suits complex coding and reasoning work.
- Lower per-token prices can raise total spend if teams increase usage without checking real costs.
Frequently Asked Questions About GPT-6 Sol and Luna
What is the difference between GPT-6 Sol and GPT-6 Luna?
Sol handles complex coding, debugging, and reasoning tasks. Luna handles high-volume, repetitive work like summarization and data extraction at a much lower per-token price.
How much do GPT-6 Sol and Luna cost?
Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 per million input tokens and $0.50 per million output tokens.
Are GPT-6 Sol and Luna available in the API yet?
Yes. Both are available now in the OpenAI API as gpt-6-sol and gpt-6-luna, alongside ChatGPT and Codex.
How does GPT-6 Sol compare to Claude Opus 5?
OpenAI’s own benchmarks show Sol matching or beating Claude Opus 5 on AutomationBench and Agents’ Last Exam. Sol does it at 9 to 60 percent of the cost per task. These figures come from OpenAI, not an independent third party.
Did GPT-6 Sol and Luna replace GPT-5.6?
Yes. GPT-6 Sol and Luna replaced GPT-5.6 Sol and Luna as OpenAI’s default models on September 22, 2026.
Should my team switch to GPT-6 Sol and Luna right away?
Test both models against your own prompts and traffic first. The lower price only helps your budget if the models still meet your accuracy bar on your actual tasks.
Final Thoughts
GPT-6 Sol and Luna matter less for their benchmark scores than for what the price cut signals. OpenAI is betting that cheaper, task-specific models drive more usage than one expensive general model. Before switching, test both models against your own prompts and real traffic patterns. Check your cache hit rate too, since that often affects your bill more than the headline price. Match the model to the task instead of defaulting to the biggest one available.
Photo by Levart_Photographer: Unsplash
Noah Nguyen is a multi-talented developer who brings a unique perspective to his craft. Initially a creative writing professor, he turned to Dev work for the ability to work remotely. He now lives in Seattle, spending time hiking and drinking craft beer with his fiancee.
