OpenAI introduces GPT-6 Sol and Luna at lower cost

OpenAI has introduced GPT-6 Sol and GPT-6 Luna, two models designed to bring the capabilities of GPT-6 Astra to everyday tasks, with API prices 50% lower than GPT-5.6 promotional pricing.

The company continues to position GPT-6 Astra as its most capable model for work that demands the highest level of reasoning. Sol and Luna occupy a more practical tier: they aim to deliver strong results in programming, automation and computer use without making every task so expensive.

More capability for the same budget

OpenAI says it applied training methods similar to Astra's to Sol and Luna, including advances in professional work, factual accuracy, programming, tool use and alignment. The difference lies in cost and speed.

In the tests shared by the company, GPT-6 Sol delivers competitive results against Claude models at a much lower cost:

• In AutomationBench, which measures business workflows across different applications, Sol with xhigh effort outperforms Claude Opus 5 at maximum effort for approximately 9% of its cost per task.
• In Agents’ Last Exam, a test of complex professional workflows, Sol reaches 56.4%, above Claude Opus 5's best published score on that evaluation, with a per-task cost 60% lower.
• In OSWorld 2.0, which evaluates computer use, Sol achieves a result similar to Claude Opus 5 with medium effort: 60.5% versus 60.3%, but at an approximately 80% lower cost.

Luna is designed for tasks that are even more price-sensitive. On AutomationBench, it improves by 5.4 percentage points over its predecessor at a per-task cost 58% lower. On DeepSWE v1.1, a software engineering test using real repositories, it reaches 66.6% at maximum effort, at a cost 93% lower than Claude Opus 5 and 96% lower than Claude Fable 5.

These figures come from OpenAI evaluations or public reports from other companies. The company warns that the tests do not necessarily represent normal use and that results may vary between its research environment and ChatGPT.

Fewer errors and clearer answers

OpenAI says Sol makes approximately half as many errors as its predecessor in an internal factual accuracy evaluation based on de-identified real conversations. Luna also improves on GPT-5.6 Luna and, at high effort levels, matches GPT-5.6 Sol at approximately one hundredth of its cost.

Both models also inherit the communication style introduced with Astra: clearer answers, less jargon, fewer unhelpful details and somewhat shorter writing, especially in technical and programming conversations.

What changes for developers

Lower prices are not the only measure aimed at reducing costs. OpenAI has also improved prompt caching, a system that allows parts of an already processed instruction to be reused instead of recalculating them from scratch.

Reads of cached inputs receive a 90% discount. According to OpenAI, GitHub has reduced by more than 50% the proportion of tokens that need to be processed again after analyzing billions of requests.

Developers can also:

• See which parts of their instructions are being reused through a cache dashboard.
• Change the reasoning level or available tools without breaking the stored context.
• Choose more precisely which parts of a prompt should be saved.

For people who use coding agents for hours at a time, the savings can be significant. OpenAI says daily token usage already amounts to more than 600 dollars for the company's median researcher and more than 7,000 dollars for the 90th percentile. When an agent works for a long time, the cost per task stops being a minor detail.

Availability

GPT-6 Sol and GPT-6 Luna are available starting today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access Luna in the desktop app.

They are not yet available in Chat, and the rollout in ChatGPT Work and Codex will be gradual. In the API, their names are gpt-6-sol and gpt-6-luna.

The central idea is simple: GPT-6 Astra is reserved for the most demanding work, while Sol and Luna aim to make advanced intelligence viable for more tasks and longer periods. The important question now is whether those savings hold up in real-world use, especially with agents that program, move between applications and need context for many hours.