Sun and Moon from OpenAI. New GPT-6 models make fewer mistakes and cost half as much

Tuesday, September 22, was a busy day in the world of artificial intelligence. Anthropic released Claude Opus 5.5 in the evening, and about an hour later, OpenAI responded. To its flagship model GPT-6 Astra, which was introduced at the beginning of the month, it added two smaller siblings: GPT-6 Sol and GPT-6 Luna. Both companies this time focused mainly on one topic – price.

Star, Sun, and Moon

The new GPT-6 series has three tiers. At the top remains Astra for the most demanding tasks. Sol is the mid-range class designed for more complex work, typically programming. Luna is the cheapest model for large volumes of simple tasks with a clear goal, such as summarizing documents, extracting data, or quick queries. According to OpenAI, both new models were created using similar methods as Astra, so some of its capabilities are also available in the cheaper variants.

Half price, and this time permanently

Developers will pay for GPT-6 Sol $2 per million input tokens and $10 per million output tokens; for the previous GPT-5.6 Sol, it was $4 and $20. Luna costs $0.10 and $0.50. An important detail: the prices for the 5.6 series were promotional from the start, while the new prices are standard, according to an OpenAI spokesperson. The company explains the price reduction by more efficient operation and better caching. Reused context, such as a long conversation, is charged with a 90% discount.

Fewer errors, shorter answers

According to OpenAI, GPT-6 Sol makes about half as many factual errors as its predecessor in an internal accuracy test. However, the test is based on conversations where users caught the previous model making mistakes, so according to the company itself, it does not represent typical usage. Both models also adopted Astra’s more economical style: less jargon, fewer unnecessary details, and somewhat shorter answers.

Battle of the benchmarks

OpenAI traditionally targets Anthropic in its benchmarks. In the DeepSWE programming test, GPT-6 Sol scores 68.8%, while Claude Fable 5 achieves 69.9%. According to OpenAI, Sol requires approximately 80% lower task costs for this. However, the comparison refers to older Claude models; the new Opus 5.5 from the same day is missing from these. In terms of list price, Sol is half the price per token compared to Opus 5.5 ($4 and $20), but each company measures differently, with different settings and costs. Who truly offers more for less will only be revealed by independent tests.

Where to find them

Both models are gradually appearing starting today in ChatGPT Work and in the Codex programming tool for Plus, Pro, Business, Enterprise, and Edu subscribers. Users of the free version and Go plan will get Luna in the desktop application. They are not yet available in the classic chat. Developers can find them in the API and GitHub Copilot. OpenAI may announce further news at the developer conference DevDay on September 29.

Do you use ChatGPT or Claude more? And does price play a role for you?

Source: OpenAI, TechCrunch, The New Stack, 9to5Google