23 September 2026
OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, and both models cost about half of what their GPT-5.6 equivalents did. GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna, aimed at high-volume work, costs $0.10 and $0.50. Both models carry a 1,050,000 token context window.
What are GPT-6 Sol and GPT-6 Luna?
GPT-6 Sol is the working model for coding, agent tasks and business workflow automation. It sits below GPT-6 Astra, which OpenAI released on 4 September 2026, but it now beats Astra on OpenAI’s own business task benchmark. GPT-6 Luna is the small, fast option for jobs you run thousands of times a day, such as pulling fields out of documents or summarising tickets.
Both models accept text and images and return text. Sol has a knowledge cutoff of 20 April 2026 and Luna of 18 May 2026. Each can write up to 128,000 tokens in one response. You control how hard the model thinks with a reasoning effort setting that runs from none through to max, so the same model can be cheap and quick or slow and thorough.
Where you can use them today:
- API model names: gpt-6-sol and gpt-6-luna
- ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Education users
- GPT-6 Luna in the desktop app for Free and Go users
- Cached input reads discounted by up to 90 percent, within a 30 minute reuse window
- A prompt caching dashboard, cache diagnostics, explicit cache breakpoints and cache prewarming
GPT-6 Sol benchmarks and pricing
Take this from the table: Sol trades a point or two of raw score for a very large drop in cost per task.
| What is measured | GPT-6 Sol | Best rival in OpenAI’s own table |
|---|---|---|
| DeepSWE v1.1 (agentic bug fixing) | 68.8% at max effort | 69.9% (Claude Fable 5) |
| OSWorld 2.0 (using a computer) | 60.5% | 60.3% (Claude Opus 5) |
| AutomationBench 1.0.6 (business tasks) | 33.2% at $0.27 per task | Beats GPT-6 Astra and Claude Opus 5 |
| Agents’ Last Exam (long agent runs) | 56.4% at max effort | Beats Claude Opus 5 at 60% lower cost |
| Price per million tokens | $2 in, $10 out | $10 in, $50 out (Claude Fable 5.1) |
| Context window | 1,050,000 tokens | 1,000,000 tokens (Claude Fable 5.1) |
These are OpenAI’s own evaluations. OpenAI states in the announcement that the competitor figures were taken from publicly available reports rather than rerun in house, and that the scores do not measure how often a model fails in normal day to day use.
GPT-6 Luna is the sharper jump. It scores 66.6 percent on DeepSWE v1.1 at max effort, which is 2.2 points behind Sol, while costing 93 percent less per task than Claude Opus 5. Against its own predecessor it gains 5.4 percentage points and cuts cost per task by 58 percent.
What OpenAI is not saying about GPT-6 Sol
The comparison table aged badly within hours. OpenAI benchmarks Sol against Claude Opus 5 and Claude Fable 5.1. Anthropic released Claude Opus 5.5 the same day at $4 and $20 per million tokens, scoring above Fable 5.1 on most of its own coding benchmarks. The cost advantage OpenAI claims over Anthropic is real but smaller than the post implies, because the Anthropic model it measures against is no longer the one you would buy.
Two other gaps are worth noting. OpenAI gives both models a 1,050,000 token window but publishes no long context benchmark to show how well either holds up at that length. And Sol Pro and Luna Pro variants appeared on OpenRouter on launch day without being mentioned in the announcement, so what separates them from the base models is not documented.
What this means
GPT-6 Luna is the release worth testing first, and it is the one most teams underrate. A support team running 40 million tokens a month through a classification or extraction step now pays roughly $4 for input where GPT-5.6 Luna charged about $8, and gets a model that is within striking distance of Sol on coding tasks. That is the kind of change that moves a pilot into production.
GPT-6 Sol is worth testing now if you already run GPT-5.6 Sol, because the price halves and the scores go up. If you are choosing fresh, run your own evaluation against Claude Opus 5.5 rather than the Opus 5 numbers in OpenAI’s table, and measure cost per finished task rather than cost per million tokens. Reasoning effort settings change token use by a wide margin, so the headline price tells you very little until you have fixed the effort level your workload actually needs. The first thing we would test is a long context retrieval job, since that is the one claim OpenAI has published no evidence for.
For more information, visit the official announcement of GPT-6 Sol and GPT-6 Luna on the OpenAI blog.
