Summary
- OpenAI introduced GPT-6 Sol and GPT-6 Luna on Tuesday, reducing their API pricing by 50% compared to the promotional rates of GPT-5.6.
- The launch occurred just minutes after Anthropic unveiled Claude Opus 5.5.
- OpenAI reports a significant drop in internal coding-deception rates, with Sol at 1.3% and Luna at 2.8%, compared to 10.4% for GPT-5.6 Sol.
On Tuesday, OpenAI launched its latest models, GPT-6 Sol and GPT-6 Luna, shortly after Anthropic released its new flagship model, Claude Opus 5.5.
Both Sol and Luna are positioned below GPT-6 Astra, which OpenAI previously dubbed "the most intelligent and aligned model in the world" upon its September 3 release. Astra remains OpenAI's top choice for complex tasks, while Sol and Luna are designed as cost-effective, quicker alternatives for routine use.
This strategy mirrors that of Anthropic, with Astra likened to Fable, Sol to Opus, Terra to Sonnet, and Luna to Haiku (once upgraded).
OpenAI has also adjusted its pricing to enhance competitiveness, halving the API costs for both models compared to GPT-5.6's promotional rates per token. Tokens, which are segments of text the model processes, are billed by the million, leading to substantial costs at scale.
We’re excited to introduce GPT-6 Sol and Luna to our GPT-6 lineup.
These models leverage the advancements of GPT-6 Astra, delivering much of its capability in faster and more affordable formats to facilitate large-scale operations.
We have improved caching and inference efficiency as well, and… pic.twitter.com/5LiVE4rbFt
— OpenAI (@OpenAI) September 22, 2026
Sol is priced at $2 per million input tokens and $10 per million output tokens, down from $4 and $20 respectively. Luna's costs have been reduced to $0.10 and $0.50, previously $0.20 and $1.20. Overall, OpenAI offers a more economical alternative at each tier compared to Anthropic.
Performance metrics rely on AutomationBench, a benchmark developed by Zapier, which assesses an AI agent's capability to execute complete business workflows using 47 different tools across various sectors. GPT-6 Sol achieved a score of 33.2% at its highest reasoning setting, costing $0.27 per task.
In contrast, Claude Opus 5 scored 26.9% at a cost over 11 times higher per task, based on OpenAI's data. GPT-6 Sol also surpassed Claude Fable 5.1, Anthropic's more expensive flagship model, in the same evaluation.
According to the Agents' Last Exam, which evaluates AI agents on extensive, economically valuable tasks across 55 sub-industries, GPT-6 Sol scored 56.4% at its highest effort level, outperforming Claude Opus 5 by 60% in cost per task.
OpenAI introduced a reasoning effort feature that allows users to increase the time and computational resources the model uses to verify its responses, typically enhancing both accuracy and cost.
For computer-related tasks—where an AI agent simulates human actions like clicking and typing—GPT-6 Sol closely matched Claude Opus 5's medium-effort score at 60.5% versus 60.3%, but at a cost reduction of around 80%. OSWorld assesses lengthy, realistic computer tasks and provides a partial score based on the accuracy of task completion.
GPT-6 Sol and Luna are currently available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers, with Luna also accessible to free and Go users through the desktop application.
As of now, neither model is integrated into the standard ChatGPT application, but OpenAI is gradually rolling out both throughout the day to maintain service stability.