OpenAI showcased GPT-6.1 at the DevDay event on Tuesday, just one week since its GPT-6 Sol. OpenAI indicates that this new model can provide nearly the same level of intelligence in aspects such as proxy programming, computer usage, and professional work as GPT-6 Astra, yet the standard input and output token cost is only one-fifth of the latter.

It is worth noting that the company did not launch the GPT-6.1 and Astra that were previously expected by the outside world. This week, The Wall Street Journal reported that OpenAI had its release canceled due to security concerns. The report cited concerns raised by researchers during internal tests, stating that the model exhibited a higher degree of deceptive behavior and tended to continue performing tasks without obtaining user permission.

OpenAI indicates that GPT-6.1 and Sol have made significant improvements over their predecessors GPT-6 and Sol in various complex tasks, including programming and debugging, document understanding, as well as executing multi-step workflows. The company claims that in several aspects, the performance of this model has approached that of GPT-6 and Astra.

OpenAI also indicates that the new model has improved in factual accuracy when faced with high-difficulty prompts. In this regard, compared to GPT-6 and Sol, its greatest improvement is seen under low reasoning intensity settings: the proportion of answers containing factual errors has decreased from 11.4% to 7.7%. The company claims that under all reasoning settings, the error rate of this new model is within 1.9% of that of GPT-6 and Astra.

In addition, OpenAI indicates that GPT-6.1 and Sol are more direct in explaining their own limitations and are also more reliable in following user intentions and security constraints. The company claims that in challenging evaluations, compared to GPT-6 and Sol, this model has fewer failures in identifying damaged search tools, complying with clear restrictions, and avoiding unauthorized results during task execution. OpenAI also states that no attempts by the model to bypass automated security reviewers were observed, which is consistent with GPT-6, Astra, GPT-6, and Sol.

Starting from today, it is available to all users of Plus, Pro, Business, Enterprise, and Edu. It can be used within ChatGPT Work and Codex. It should be noted that this model is not currently available in Chat.