OpenAI gives its Open Responses API spec to the AAIF

On October 22, OpenAI handed Open Responses, its specification for an LLM API that works across providers, to the Agentic AI Foundation (AAIF). The AAIF is part of the Linux Foundation and already hosts MCP and AGENTS.md. Two gateways already run the spec: an OpenAI spokesperson said Vercel AI Gateway and Databricks Model Serving implement it.

Three other projects joined the foundation on the same day, The New Stack reported: OpenSandbox, Anthropic’s Agent Skills and HuggingFace’s OpenEv framework for agent environments. They sit alongside the A2A protocol, goose, agentgateway and Agent Router. The foundation launched in December 2025 with three projects: MCP, goose and AGENTS.md.

OpenAI first announced Open Responses on January 15. At the time the team argued that nearly every LLM provider offers the same building blocks, such as messages, tool calls and function calling, but that each API represents them differently. The spec defines those primitives once, so developers who switch models don’t have to translate between vendor formats. OpenAI’s announcement post summed up the pitch as “One spec, many providers: Define inputs and outputs once and run them across OpenAI, Anthropic, Gemini, or local models.” The post also promised simpler evaluation and routing through a shared schema, plus a blueprint providers can follow to expose their APIs in a common format.

The basic unit in the spec is the item, per the project’s GitHub repository. An item is one piece of model input or output, a tool invocation or a reasoning state, and it has a type such as message, function_call or reasoning. Providers can emit custom item types the spec doesn’t define. The spec is built around an agentic loop: the model emits tool calls, gets the results back and continues. Streaming is defined as semantic events rather than raw text deltas, so a single client can read any provider that follows the spec.

The spec also separates two kinds of tools, InfoQ explained. Internal tools run on the provider’s infrastructure, so the model drives the loop itself and returns a final result in a single API round-trip. External tools run in the developer’s own code: the provider pauses to request a tool call, the application sends back the output, and the model continues. The reasoning item leaves disclosure to the provider, which can return raw reasoning content, protected content or only summaries.

The schema is written in TypeSpec and generated into an OpenAPI document. Dated releases under public/openapi/<YYYY-MM-DD>/ are immutable, and the repo keeps a specification changelog. To check an endpoint against the spec, there is an interactive compliance tester on the docs site and a CLI runner, bin/compliance-test.ts. Each HTTP request, response body included, has a 30-second deadline, and a timeout fails only that test while the rest keep running.

The list of supporters is already long. AWS supports the spec in Bedrock, and so do Databricks, LM Studio, Ollama, Hugging Face, Nvidia, OpenRouter, Red Hat and Vercel, along with the open source projects vLLM and OGC (formerly Llama Stack).

Anthropic is not on that list. It keeps its own Messages API for access to its models, even though it is an AAIF member and contributed Agent Skills on the same day. Google has its own interface too, the Interactions API. Those are the two providers OpenAI’s January post named, next to OpenAI and local models, as places the same inputs and outputs could run.

Open Responses is open about where it comes from. Its contributor guide says the core schema files are copied from OpenAI’s API. OpenAI’s earlier Chat Completions API became an early de facto standard that other providers modelled their APIs on. The New Stack noted that Anthropic’s Messages API has since had a similar effect, largely because of Claude Code’s success: DeepSeek, Moonshot, Z.ai, MiniMax and others now expose Anthropic-compatible endpoints as well.

The OpenAI lineage drew comment when the spec was first announced. InfoQ quoted commenter Rituj Pramanik: “Building an \”open\” standard on top of OpenAI’s API is slightly ironic, but practical.” In the OpenAI Developer Community thread on the announcement, user _j wrote that the spec “imposes, embraces, and extends OpenAI’s limitations to ensure an uncompetitive surface if employed,” and said it cut streaming down to 23 event types.

For OpenAI, the donation follows the one it made when the foundation launched. “When we co-founded AAIF, we contributed AGENTS.md, a way to give coding agents project-specific instructions,” the spokesperson told The New Stack. “Open Responses addresses another part of building agents by giving developers a shared format for working with models across endpoints that support it, so they have fewer provider-specific integrations to maintain.” According to the spokesperson, the AAIF “will provide a neutral home for developers and providers to continue developing the specification together.” Technical discussions will go to a technical steering committee that at first will be made up of the project maintainers.