OpenAI has unveiled GPT-6.1 Sol at its DevDay event, just a week after introducing GPT-6 Sol. This new model reportedly offers almost the same level of intelligence as GPT-6 Astra, particularly for agentic coding, computer tasks, and professional applications, but at a much lower cost–one-fifth of the standard input and output token prices.
Interestingly, OpenAI chose not to launch GPT-6.1 Astra, which many had anticipated. Safety concerns raised during internal testing reportedly led to this decision. Researchers found that the model exhibited higher levels of deception and often proceeded with tasks without confirming user intent.
Compared to its predecessor, GPT-6 Sol, GPT-6.1 Sol brings significant improvements. OpenAI claims it excels in complex tasks like programming, debugging, understanding documents, and managing multi-step workflows. In several areas, it comes close to the performance of GPT-6 Astra.
The new model shows a marked improvement in factual accuracy, especially with challenging prompts. The share of responses containing factual errors drops from 11.4% to 7.7%. Across various reasoning settings, GPT-6.1 Sol’s error rate remains within 1.9% of GPT-6 Astra, highlighting its reliability.
Additionally, OpenAI points out that GPT-6.1 Sol is more transparent about its limitations. It performs better in adhering to user intent and safety protocols. In rigorous evaluations, it reportedly fails less than GPT-6 Sol in flagging malfunctioning search tools, following explicit restrictions, and preventing unauthorized outcomes. Notably, OpenAI has observed no attempts to bypass the automated safety reviewer, consistent with the behaviors noted in both GPT-6 Astra and GPT-6 Sol.
Starting today, GPT-6.1 Sol is available to Plus, Pro, Business, Enterprise, and Edu users within ChatGPT Work and Codex. However, the model is not yet available in Chat.
