The Build Hour session introduces the new Responses API, which enhances developer capabilities by supporting multi-step reasoning, tool integration, stateful conversations, and multimodal inputs, enabling more advanced and interactive AI-driven applications. Through live demos and migration guides, the session showcases how the API simplifies complex workflows and previews upcoming tools like Agent Kit, while addressing common developer questions and best practices.
In this Build Hour session, Christine from the startup marketing team and Steve, an engineer on the API team, introduce the new Responses API, designed to empower developers to build more advanced and agentic applications using OpenAI’s latest models. They explain that the Responses API is a significant evolution from the previous chat completions API, offering enhanced capabilities such as multi-step reasoning, tool integration, and state management. The session includes live demos, migration guides, and a preview of upcoming tools like Agent Kit, aiming to simplify the development of complex AI-driven workflows.
Steve provides a historical overview, explaining how OpenAI’s API evolved from the original completions API in 2020, which was designed for simple text continuation, to the chat completions API introduced with GPT-3.5 Turbo in 2022, which was optimized for conversational interactions. However, with the release of more advanced, multimodal, and agentic models like GPT-4 and GPT-5, there was a need for a more flexible API. The Responses API addresses this by enabling multiple model samplings per request, supporting complex workflows such as code execution, tool use, and long multi-turn interactions, all within a single API call.
A key innovation of the Responses API is its use of “items” instead of just messages, allowing the model to output different types of content such as reasoning steps, function calls, or tool invocations. This design makes it easier for developers to handle diverse outputs and maintain state across interactions. The API also supports stateful conversations by preserving the chain of thought between requests, which improves performance and reduces costs compared to the chat completions API. Additionally, the Responses API is built to handle multimodal inputs like images and PDFs, and it features a redesigned streaming system that emits strongly typed events for easier integration.
The session includes practical demonstrations, such as migrating a simple chat application from the chat completions API to the Responses API using an automated migration pack. Steve also showcases a custom-built game simulating a day in the life of an OpenAI engineer, highlighting how the Responses API can manage agentic characters with access to various tools like web search, image generation, and task management via MCP (Model Control Protocol). These demos illustrate how the API enables rich, interactive experiences by allowing agents to reason, call external tools, and provide real-time feedback through streaming events.
In the Q&A segment, Steve and Christine address common questions about prompt engineering, performance differences, context management, and best practices for using the Responses API. They emphasize the benefits of few-shot prompting to reduce hallucinations, the improved efficiency of the Responses API in multi-turn workflows, and the various methods for preserving conversation context, including the use of conversation objects and encrypted content. They also discuss the MCP tool integration, which allows agents to interact with external services dynamically. The session concludes with a preview of upcoming Build Hours focused on Agent Kit, reinforcement fine-tuning, and memory patterns, encouraging developers to continue exploring and building with OpenAI’s evolving platform.