Open Responses: Unified LLM interface instead of adapter chaos

Open Responses, an open-source standard for a vendor-independent JSON API, has been released, enabling language models to communicate with clients. It is based on the Responses API and represents another building block of the company's agentic AI ecosystem.

Together with the Responses API, OpenAI last year introduced integrated tools and a Software Development Kit (ADK) to create custom AI agents. The Responses API combines OpenAI's Chat Completions API and Assistants API and can act independently, rather than being limited to delivering answers. Until now, however, a separate client adapter was required for each LLM. Open Responses is now intended to standardize this.

In the coming months, OpenAI plans to further develop the open standard together with the community and inference solution providers. To this end, OpenAI has enlisted Hugging Face, Nvidia, LLM, LM Studio, Ollama, OpenRouter, and Vercel as launch partners.

Mozilla already has a similar project underway with any-llm. The Python package provides a unified API for many LLMs, saving developers the effort of maintaining a separate adapter for each individual LLM.

Easy transition from Responses API to Open Responses

To improve AI interoperability, Open Responses defines a “common schema and tool layer to standardize the invocation of language models, the streaming of results, and the composition of agent-based workflows.” This is intended to work independently of the provider.

According to OpenAI, developers already using the Responses API can switch to the new format with little effort. The changes primarily concern agentic reasoning. In addition to encrypted_content (provider-specific protected content) and summary (data cleaned from reasoning traces), and the new content (reasoning traces), three input parameters are now available. The latter allows reasoning traces to be made accessible via the API, making it easier to switch providers.

In the case of AI Delta Learning, the function calls for Open Response look like this compared to Responses API:

// Open weight models stream raw reasoning
event: response.reasoning.delta
data: { "delta": "User asked: 'Where should I eat...' Step 1: Parse location...", ... }

// Models with encrypted reasoning send summaries, or sent as a convenience by Open Weight models
event: response.reasoning_summary_text.delta
data: { "delta": "Determined user wants restaurant recommendations", ... }

Those who want to delve deeper into the agentic open-source model can find a technical description on its website.

(mki)