OpenAI officially launched its new Responses API on Tuesday, an innovative tool designed to provide stronger support for developers and enterprises to leverage OpenAI's models and frameworks to build AI agents that can independently complete complex tasks. This new tool will gradually replace the existing assistant API, which is expected to be officially discontinued in the first half of 2026.
The core function of the response API is to allow enterprises to develop highly customized AI agents that can perform a variety of tasks, including network searches, company file scanning, and website browsing. Its functions are similar to OpenAI's Operator products. Developers can use the same AI models as ChatGPT Search, namely gpt-40 search and gpt-40 mini search, which not only find answers to questions on the Internet, but also reference relevant sources when generating replies to ensure the accuracy and reliability of information.
"While demonstrating AI agents are relatively easy, scaled up and letting users use them frequently is a huge challenge," Olivier Goldmond, product director of OpenAI, said in an interview. He further pointed out that OpenAI's goal is to narrow the gap between demonstration and practical applications within this year, and emphasized that "agents are one of the most promising application areas of artificial intelligence."
In addition to the response API, OpenAI has released an open source toolkit called the Agents SDK, providing developers with a range of free tools to seamlessly integrate AI models with internal systems, implement necessary security measures, and monitor the activity of AI agents for debugging and optimization.
Although OpenAI claims its search model is excellent in factual accuracy (gpt-40 search scored 90% on the benchmark and gpt-40 mini search scored 88%), the company also admitted that there are still some limitations in these tools. For example, web search does not completely solve the AI hallucination problem, and computers use agents (CUA) models to achieve a completely reliable level in automated tasks on operating systems.
OpenAI official blog post: https://openai.com/index/new-tools-for-building-agents/