OpenAI takes on rivals with new Responses API, Agents SDK

OpenAI’s new Responses API and an upgraded Agents SDK will help enterprises more easily build agents with advanced reasoning and multimodal capabilities, the company said Wednesday.

The new tools may help it fend off challenges from rivals such as Anthropic or up-and-coming Chinese competitors including DeepSeek and Butterfly Effect, the developer of Manus, looking to capture a chunk of the agentic automation market.

OpenAI has upgraded its Agents SDK with new integrated observability tools to trace and inspect agent workflow execution

The Responses API combines capabilities of the existing Chat Completions and Assistants API, and will become the de facto API that enterprises use to build agents to handle complex tasks.

OpenAI said that repackaging the capabilities in this way will help developers incorporate its built-in tools into their apps, without the complexity of integrating multiple APIs or external vendors.

“Developers feel like they’re having to cobble together different low level APIs from different sources. It’s difficult, it’s slow, it often feels brittle,” Kevin Weil, chief product officer at OpenAI, explained in a webcast with reporters.

Capabilities packaged in the Responses API

For now, the LLM provider is packaging three capabilities — web search, file search, and computer use — to help developers connect models to the real world and make them more useful in completing tasks.

The web search tool is the same one that powers ChatGPT Search, underpinned by a fine-tuned GPT-4o and GPT-4o mini models, said Nikunj Handa, an engineer in OpenAI’s product team, in the same webcast.

The file search tool, initially introduced last year as part of the Assistants API to help developers perform RAG on documents, now includes metadata filtering so developers can search on file attributes and a direct search endpoint that can comb data stores without queries being filtered through the AI model, explained Steve Coffey, an engineer with OpenAI, during the webcast.

The third tool packaged inside the Responses API is the computer use tool, which uses the same operator model found in ChatGPT.

“The computer use tool is operator in the API, but it allows you to control the computers that you are operating. So, this could be a virtual machine or a legacy application that just has a graphical user interface and developers have no API access to it,” Handa explained during the webcast.

Rival Anthropic introduced such a computer-use capability accessible through an API in its Claude 3.5 Sonnet LLM last October. This can read and interpret what’s on the computer’s display, type text, move the cursor, click buttons, and switch between windows and applications.

There are significant differences between the two companies’ approaches.

Moor Insights and Strategy principal analyst Jason Andersen noted that OpenAI’s system is screenshot based, whereas Anthropic can work with command outputs from tools.

In the same vein, Forrester vice president and principal analyst Charlie Dai pointed out that the two LLM providers have different design philosophies, safety considerations, and ecosystem integration, which may lead to variations in how abilities such as computer use are implemented and constrained.

“OpenAI’s models might feel more versatile and widely applicable, while Anthropic’s models might prioritize safety and alignment in their approach to computer use based on their claims, such as ‘We’ve trained the model to resist these prompt injections and have added an extra layer of defense,’” Dai said.

Andersen said one advantage of the new API is that it opens up an avenue for developers to automate tasks without a major migration effort. But he warned that the cost of this approach may add up over time as users and task complexity scale.

The Responses API is currently available and is not charged separately, which means that enterprises pay for the tokens and tools at OpenAI’s standard rates.

Chat Completions: not dead yet

OpenAI said it will continue to support Chat Completions, its most widely adopted API, and add new models to it, even though it is now subsumed within the Responses API. If developers don’t use built-in tools they can continue to use the Chat Completions API, but if they require built-in tools, especially for newer integrations, the Responses API should be the preferred choice, the company said.

As for the Assistants API, OpenAI’s plan is to include every tool and feature into it along with the Responses API, before phasing it out in mid-2026. When it formally announces the API’s deprecation, it said, it “will provide a clear migration guide from the Assistants API to the Responses API that allows developers to preserve all their data and migrate their applications.”

What is OpenAI’s Agents SDK?

The open source Agents software development kit (SDK) is an upgraded version of Swarm⁠, an experimental SDK that OpenAI released last year to help developers orchestrate agentic workflows. Despite Swarm’s experimental status, OpenAI said several enterprises have already adopted it.

Some of the improvements that accompany its rebranding as Agents SDK include new agents, handoffs between agents, guardrails, and observability tools to debug agents and trace their performance.

Andersen said the SDK is a “big deal” since platforms such as Amazon Bedrock and Google’s Vertex AI are rapidly expanding into providing workflow and agent-to-agent collaboration support.

“But, it’s also significant since OpenAI has been a proponent of very large broadly trained models. The idea of collaborative agents suggests that OpenAI is also open to the idea of small, focused models for specific tasks working with other models,” he said.

The Agents SDK, according to OpenAI, works with the Responses API and Chat Completions API.

The SDK will also work with models from other providers, as long as they provide a Chat Completions-style API endpoint, the LLM provider wrote, adding that developers can immediately integrate it into their Python codebases, with Node.js support coming soon.