In addition to the high-end GPT-6 Astra, OpenAI has introduced two more affordable models: GPT-6 Sol and GPT-6 Luna.
The new models, the company states, are designed for professional work, coding, automation and computer-use tasks in which users want powerful AI without the price of Astra.
GPT-6 Sol offers a $2 price per million input and $10 price per million output tokens via API. The cost of using GPT-6 Luna is $0.10 per million tokens for input and $0.50 per million tokens for output. The pricing of both models is half that of the promotional pricing of OpenAI’s GPT-5.6 models, according to the company.
Astra is still OpenAI’s most powerful model for tough tasks. Sol and Luna, however, target the low end of the cost-intelligence curve, enabling users to do more work within the same cost range.
OpenAI claims that the new models have been trained in a similar fashion to Astra. These improvements in reasoning, factual reliability, coding, computer use and alignment are advantageous to the company, the company said.
The pricing changes are accompanied by enhancements to prompt caching. According to OpenAI, developers can get up to 90% off cached input-token reads, potentially lowering expenses for applications that frequently reuse extensive amounts of context.
The company mentioned a number of internal and external evaluations. OpenAI claims that GPT-6 Sol, at its highest reasoning setting, scored 33.2% at approximately $0.27 per task on AutomationBench. That surpassed the published results of some of the other models that were competing in the same test, OpenAI said, though the scores are based on OpenAI’s own test methodology.
Coding is a big emphasis too. GPT-6 Sol achieved a score of 68.8% on DeepSWE, which consists of complex software-engineering tasks in real codebases. At max effort, GPT-6 Luna achieved a score of 66.6%.
On the offline variant of OSWorld 2.0, GPT-6 Sol achieved a score of 60.5% for computer-use tasks, OpenAI said. That was near Claude Opus 5 at medium effort in the company’s comparison and much cheaper per task based on OpenAI’s calculations.
In its internal testing, OpenAI also claims that Sol and Luna do not make as many factual errors as the GPT-5.6 versions that came before them. The company warns that the evaluation is based on conversations where users had previously reported errors, and thus the results are not representative of typical everyday usage.
The two models are “gpt-6-sol” and “gpt-6-luna”, which are accessible via the OpenAI API. They are also being rolled out to ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users.
GPT-6 Luna is available for Free and Go users via the desktop app. The models are not currently available in the standard ChatGPT interface, and OpenAI will roll them out gradually, it said.
The launch provides OpenAI with a broader array of models in the same generation. Astra targets the most challenging workloads and Sol and Luna are intended to make high-level AI more cost-effective for general workloads and large-scale applications.
The most significant difference for developers might be the reduced operating cost, and not a brand new capability class. OpenAI hopes that these improvements, which include lower inference costs, better caching, and model efficiency, will push customers to adopt GPT-6 for a broader range of use cases and on a larger scale.
