OpenAI Launches GPT-6 Sol and Luna, With API Prices Cut by Half or More

OpenAI unveiled two new models, GPT-6 Sol and GPT-6 Luna, on September 22 local time. Positioned as faster and cheaper siblings to the earlier GPT-6 Astra, the pair improve on task execution, factual accuracy, coding, computer use (the ability to operate a screen and apps directly to complete tasks), response style, and alignment (the work of tuning a model to behave safely and as intended), according to the company.

Prices cut in half

For developers, the most striking change is API pricing. Sol is priced at $2 per million input tokens and $10 per million output tokens, half the $4 and $20 charged for the previous generation GPT-5.6 Sol. The lightweight Luna comes in at $0.10 for input and $0.50 for output, down from $0.20 and $1.20 for GPT-5.6 Luna. OpenAI summarized the reduction for both models as 50% against previous promotional pricing, though Luna’s output price falling from $1.20 to $0.50 works out to roughly a 58% cut. Prompt caching has also been improved: re-reading input tokens already sitting in the cache earns a 90% discount. That is where real cost differences can add up, in agents or in-house chatbot setups that feed the same instructions over and over.

The benchmarks, and their limits

On AutomationBench, an evaluation of workplace automation, Sol (xhigh setting) scored 33.2% at $0.27 per task. For comparison, OpenAI cited Astra (low) at 30.3% and Claude Opus 5 (max) at 26.9%. Luna (high) reportedly scored 5.4 percentage points higher than its predecessor while costing 58% less per task. On Agents’ Last Exam, Sol (max) scored 56.4%, and on the DeepSWE 1.1 coding evaluation Sol (max) hit 68.8% and Luna (max) 66.6%. On the OSWorld 2.0 offline evaluation, which measures computer use, Sol (xhigh) reached 60.5%. In internal factual accuracy testing, OpenAI said Sol made roughly half as many errors as the previous model.

The company attached caveats, however. The factual accuracy evaluation is built from conversations in which errors were reported, so it cannot be read as an error rate for typical usage, and the benchmark figures were measured in research settings or via the API, so they may differ from what ChatGPT actually outputs. Competitors’ scores were taken from published reports. Both models also improved over their respective GPT-5.6 predecessors on alignment evaluations, OpenAI said, and make fewer misleading claims about coding work.

Where you can use them

Starting the day of the announcement, Plus, Pro, Business, Enterprise, and Edu users can access both models in ChatGPT Work and the coding tool Codex. Free and Go users can use Luna in the desktop app. The API model names are gpt-6-sol and gpt-6-luna. As of the announcement, though, neither model was yet available in Chat; OpenAI said the ChatGPT rollout would proceed in stages over the course of the day.

The heart of this release is not a new top-performing model but a combination: better performance within the same family at a lower price. That eases the token cost burden that has been a sticking point for Korean companies trying to integrate agents or coding automation into their day-to-day workflows. How well the models handle Korean, and what local pricing plans and the Chat rollout timeline look like, still need to be confirmed.