Developers can now access the same models as ChatGPT Search, GPT-4o search or GPT-4o Mini search. These models can Browse the web for answers to questions and cite references in their responses.
This is notable because OpenAI claims that the addition of web search capability dramatically improves factual accuracy in its AI models. OpenAI’s SimpleQA Benchmarkmeasures confabulation rate. GPT-4o Search scored 90 percent and GPT-4o Mini Search achieved 88 percent. Both significantly outperformed the larger GPT 4.5 model without search which scored 63 per cent.
The technology is still limited despite these improvements. GPT-4o still makes factual errors 10% of the time, despite the improvements in search.
Along with the Responses APIOpenAI released the open-source Agents SDK, which provides developers free tools to incorporate models with internal systems and implement safeguards. This toolkit is a follow-up to OpenAI’s earlier Swarmrelease, a framework that orchestrates multiple agents.
This is still early in the AI agent world, and things are likely to improve rapidly. The AI agent movement is still vulnerable to unrealistic claims. This was demonstrated earlier this week, when users discovered that Chinese startup Butterfly Effect’s Manus AI platform failed to deliver on its many promises.
