OpenAI Announces New Features and Tools to Easily Build High-Performance AI Agents
OpenAI has announced 'an initial set of building blocks to help developers and enterprises build useful and trustworthy agents.'
New tools for building agents | OpenAI
Over the past year, OpenAI has introduced new model capabilities, including advanced reasoning, multimodal interaction , and new safety techniques, laying the foundation for models that can handle the complex multiple tasks required to build agents.
However, customers have shared with us that turning these capabilities into production-ready agents is challenging, often requiring extensive prompt repetition and custom orchestration logic without sufficient visibility or built-in support.
To address these challenges, OpenAI has released a new set of APIs and tools specifically designed to simplify agent application development.
The Responses API is a new API primitive for building agents using OpenAI's built-in tools. It combines simple chat completion with the Assistant API's tool usage capabilities, allowing developers to solve increasingly complex tasks using multiple tools and model turns by calling a single Responses API. In the example embedded in the tweet below, for the user's task of 'finding a new book,' it executes three steps: 'using the file_search tool to get the user's preferences,' 'using the web_search tool to search the web and select a book,' and 'executing the displayBookList function to display the book cover.'
Introducing the Responses API: the new primitive of the OpenAI API.
— Atty Eleti (@athyuttamre) March 11, 2025
It is the culmination of 2 years of learnings designing the OpenAI API, and the foundation of our next chapter of building agents.
🧵Here's the story of how we designed it: pic.twitter.com/Qee6YEisFb
The Responses API allows you to use web search when using GPT-4o and GPT-4o-mini. According to the benchmark ' SimpleQA ' that measures the accuracy of LLM answers to short facts, GPT-o4 using web search recorded 90% and GPT-o4-mini recorded 88%, which is higher than GPT-4.5's 63%.
This web search tool is available to any developer as a preview version of the Responses API. Direct access to the fine-tuned model is also provided by the Chat Completions API through 'gpt-4o-search-preview' and 'gpt-4o-mini-search-preview'. Pricing is $30 per 1000 queries for GPT-o4 and $25 per 1000 queries for GTP-o4-mini.
In addition, an improved file search tool makes it easier to retrieve relevant information from large volumes of documents. The tool is available through the Responses API and costs $2.50 (about 370 yen) per 1,000 queries. File storage is $0.10 (about 15 yen) per day and per GB, with the first 1 GB available for free.
Additionally, computational tools are now available for building agents that can interact with computers: Research Preview models achieved a 38.1% success rate on OSWorld , a benchmark for multimodal agents, as well as 58.1% on WebArena and 87% on WebVoyager .
The computational usage tool is available as a research preview model on the Responses API to Tier 3 to Tier 5 users on higher-priced plans, with inputs priced at $3 per million tokens and outputs priced at $12 per million tokens.
In addition, an Agents SDK is also provided.
OpenAI's goal is to 'provide developers with a seamless platform experience to build agents that can assist with a wide range of tasks across any industry.'
Related Posts:
in Software, Posted by logc_nt