Open Responses is a latest and open inference standard. Initiated by OpenAI, built by the open source AI community, and backed by the Hugging Face ecosystem, Open Responses is based on the Responses API and is designed for the future of Agents. In this blog post, we’ll look at how Open Responses works and why the open source community should use Open Responses.

It has been revealed that the era of the chatbot is long gone, and agents dominate inference workloads. Developers are shifting toward autonomous systems that reason, plan, and act over long-time horizons. Despite this shift, much of the ecosystem still uses the Chat Completion format, which was designed for turn-based conversations and falls short for agentic use cases. The Responses format was designed to address these limitations, but it is closed and not as widely adopted. The Chat Completion format is still the de facto standard despite the alternatives.

In a recent development, this mismatch between the agentic workflow requirements and entrenched interfaces motivates the need for an open inference standard. Over the coming months, we will collaborate with the community and inference providers to implement and adapt Open Responses to a shared format, practically capable of replacing chat completions.

Open Responses builds on the direction OpenAI has set with their Responses API introduced in March 2025, which superseded the existing Completion and Assistants APIs with a consistent way to:.

Open Responses extends and open-sources the Responses API, making it more accessible for builders and routing providers to interoperate and collaborate on shared interests.

It has been revealed that we’ll briefly explore the core changes that impact most community members. If you want to deep dive into the specification, check out the Open Responses documentation.

Recent findings suggest that client requests to Open Responses are similar to the existing Responses API. Below we demonstrate a request to the Open Responses API using curl. We're calling a proxy endpoint that routes to Inference Providers using the Open Responses API schema.

Industry observers note that clients that already support the Responses API can migrate to Open Responses with relatively little effort. The main changes involve how reasoning content is exposed:.

Recent findings suggest that openAI models used to only expose summary and encrypted_content. With Open Responses, providers may expose their raw reasoning via the API. Clients migrating from providers that previously returned only summaries and encrypted content will now have the opportunity to receive and handle raw reasoning streams when supported by their chosen provider.

Industry observers note that for Model Providers, implementing the changes for Open Responses should be straightforward if they already adhere to the Responses API specification. For Routers, there is now the opportunity to standardize on a consistent endpoint and support configuration options for customization where needed.

Over time, as Providers continue to innovate, certain features will become standardized in the base specification.

In summary, migrating to Open Responses will make the inference experience more consistent and improve quality as undocumented extensions, interpretations, and workarounds of the legacy Completions API are normalized in Open Responses.

You can see how to stream reasoning chunks below.

Here’s the difference between getting an Open Response and using OpenAI Responses for reasoning deltas:.

Open Responses distinguishes between “Model Providers” — those who provide inference — and “Routers” — intermediaries who orchestrate between multiple providers.

Clients can now specify a Provider along with provider-specific API options when making requests, allowing intermediary Routers to orchestrate requests between upstream providers.

Recent findings suggest that open Responses natively supports two categories of tools: internal and external. Externally hosted tools are implemented outside the model provider’s system. For example, client side functions to be executed, or MCP servers. Internally hosted tools are within the model provider’s system. For example, OpenAI’s file search or Google Drive integration. The model calls, executes, and retrieves results entirely within the provider's infrastructure, requiring no developer intervention.

Open Responses formalizes the agentic loop which is usually made up of a repeating cycle of reasoning, tool invocation, and response generation that enables models to autonomously complete multi-step tasks.

image source: openresponses.org.

For internally-hosted tools, the provider manages the entire loop; executing tools, returning results to the model, and streaming output. This means that multi-step workflows like "search documents, summarize findings, then draft an email" use a single request.

Clients control loop behavior via max_tool_calls to cap iterations and tool_choice to constrain which tools are invocable:.

The response contains all intermediate items: tool calls, results, reasoning.

In a recent development, open Responses extends and improves the Responses API, providing richer and more detailed content definitions, compatibility, and deployment options. It also provides a standard way to execute sub-agent loops during primary inference calls, opening up powerful capabilities for AI Applications. We are looking forward to working with the Open Responses team and the community at large on future development of the specification.

You can try Open Responses with Hugging Face Inference Providers today. We have an early access version available for use on Hugging Face Spaces - try it with your Client and Open Responses Compliance Tool today!

Does this mean that the next step for local llm endpoint providers (like vLLM) is to support hosted tools?

Yes - I think we will see a lot more of this pattern especially for Agents offloading work to sub-agent Tool Loops via Open Responses.

Are you going to open source the code of the API?

I hope this doesn't normalize the obscuring of raw output.

According to recent reports, maybe there is a use for this. We donot need an "agent loop" insider the vendor boxes.What we definitely need is a good omni-adopted API for basic inference without anything more than general custom tool calling.

Only putting in inference paramters and that's that. I suppose that is the reason why the community stayed with the Completions API yet, right?

Agent loops are supposed to be implemented in the agent system not in the LLM vendors system.They can simply add their proprietary Agent loop thingy as an "internal" tool.

However I would highly urge the community to consider the strategic implications for further evolution of these tools. There is the danger that specific vendor based models might be highly optimized to only provide good responses on their internal tools / subagent loops. Of course they have an incentive to control the whole workflow to secure their marketshare.

In a recent development, opposed to that: when everyone is using the APIs in their custom designed loop-procedure (or some from open source agent orchestration tools) the decision about what models to use is somewhat dependent on the basis of how well theses behave with external tool calling. This is rather the situation we want to be in.

Maybe there is a use for this. We donot need an "agent loop" insider the vendor boxes.What we definitely need is a good omni-adopted API for basic inference without anything more than general custom tool calling.

Industry observers note that only putting in inference paramters and that's that. I suppose that is the reason why the community stayed with the Completions API yet, right?

Agent loops are supposed to be implemented in the agent system not in the LLM vendors system.They can simply add their proprietary Agent loop thingy as an "internal" tool.

It has been revealed that however I would highly urge the community to consider the strategic implications for further evolution of these tools. There is the danger that specific vendor based models might be highly optimized to only provide good responses on their internal tools / subagent loops. Of course they have an incentive to control the whole workflow to secure their marketshare.

According to recent reports, opposed to that: when everyone is using the APIs in their custom designed loop-procedure (or some from open source agent orchestration tools) the decision about what models to use is somewhat dependent on the basis of how well theses behave with external tool calling. This is rather the situation we want to be in.

→ Read original article