OpenAI hands its Open Responses spec to the Agentic AI Foundation to standardize LLM APIs
On October 22, 2026, OpenAI transferred Open Responses, its specification for a common LLM API, to the Agentic AI Foundation, hosted by the Linux Foundation, alongside OpenSandbox and OpenEv. For teams juggling multiple model providers, it is a concrete lever for shrinking the vendor-specific integration surface.
October 22, 2026. OpenAI announced the transfer of Open Responses, its specification for a common LLM API, to the Agentic AI Foundation (AAIF), a foundation hosted by the Linux Foundation. October 22, 2026. Two other projects join it in the move: OpenSandbox and Hugging Face’s OpenEv agent-environment framework, along with Anthropic’s Agent Skills. October 22, 2026. AWS, Databricks, Nvidia, Red Hat, Vercel, and vLLM are already among the declared supporters. Why it matters: model API fragmentation is being settled from the bottom up, and the integration layer of your applications is what stands to benefit. For the first time, the cost of model-API fragmentation is being shifted away from the developer.
A problem every model integrator knows
The starting point is simple. Every LLM provider now offers the same building blocks — messages, tool calls, function calling — but each represents them differently in its API. A team that switches models must translate its calls between incompatible idioms and maintain as many integration layers as it has vendors. Open Responses targets exactly that point: a specification that standardizes these primitives so developers no longer rewrite the plumbing on every model change.
The spec’s lineage is transparent. Its contributor guide notes that its core schema files are copied from OpenAI’s API. That is consistent with the market’s history: OpenAI’s old Chat Completions API served as the de facto standard others copied, before Anthropic’s Messages API — carried by the success of Claude Code — established itself as a reference in turn, adopted by DeepSeek, Moonshot, Z.ai, and MiniMax. The market has split into two dominant dialects. That is exactly the fragmentation a neutral specification is meant to repair.
What the spec standardizes, and what it leaves out
Open Responses does not claim to unify the models themselves, only the calling layer. The primitives it targets — messages, tool calls, function calling, response formats — are the ones every provider reinvents in its own way. The spec describes a single API for expressing them, which lets a gateway or platform expose multiple models behind one contract. That is precisely what Vercel AI Gateway and Databricks Model Serving, today’s two production implementations, already do.
What it leaves out is just as important. The spec does not cover a model’s specific capabilities — certain function calling modes, image formats, streaming variants — nor the quality guarantees of an implementation. A provider can announce support for Open Responses while subtly degrading the primitives it exposes. Standardization reduces the plumbing; it does not remove the need for per-provider integration tests.
It is also worth understanding the political gesture. By handing over the spec, OpenAI gives up exclusive control of a standard that carries its DNA — the core schemas are copied from its API. It is a bet: a widely adopted neutral spec beats a proprietary standard that Anthropic and Google will ignore. The foundation, for its part, gains another project in a portfolio — MCP, A2A, AGENTS.md — that already covers most of the agent stack. The signal is clear: agent interoperability is now being built at the Linux Foundation, not in the vendors’ labs.
For a technical decision-maker, the choice is simple to state. Adopting Open Responses is a bet that the spec will become the market’s common denominator — a reasonable bet given the supporters (AWS, Nvidia, Red Hat) and the two implementations already in production. Not adopting it means keeping the freedom to optimize for one provider, at the price of a dependency you will have to own at the next switch. In between, the middle path — a gateway that speaks Open Responses on the front and keeps in-house adapters behind it — is what the most advanced teams already choose.
A neutral home, with support already broad
The choice of the AAIF matters. The foundation, launched in December 2025 within the Linux Foundation, already hosts MCP, A2A, AGENTS.md (also from OpenAI), the open source agent goose, agentgateway, and Agent Router. A technical steering committee, initially made up of project maintainers, will drive the technical discussions. This “neutral home” is generally why companies donate projects to a foundation: it turns a vendor-controlled de facto standard into a collectively governed commons.
The list of supporters is already long: AWS (in Bedrock), Databricks, LM Studio, Ollama, Hugging Face, Nvidia, OpenRouter, Red Hat, Vercel, plus the open source projects vLLM and OGC (formerly Llama Stack). Two implementations are already in production — Vercel AI Gateway and Databricks Model Serving — which gives the specification a reality beyond paper.
The absentees are just as instructive. Anthropic does not adopt the spec and keeps its Messages API; a member of the AAIF, the company could shift over time, but nothing obliges it. Google, for its part, defends its Interactions API. The standardization battle is therefore not won: it pits three dialects against each other — OpenAI’s, Anthropic’s, and Google’s — and donating Open Responses removes OpenAI’s exclusive control over its own spec without guaranteeing the other two adopt it.
What this changes for a team building agents
For an engineer assembling multi-model applications, the promise is concrete. A common specification means fewer provider-specific integrations to maintain and a smoother switch between vendors, provided vendors implement the spec faithfully. OpenAI’s statement is direct on this point: “Open Responses gives developers a shared format for working with models across endpoints that support it, so they have fewer provider-specific integrations to maintain.”
Caution remains warranted. A neutral spec does not erase each model’s particularities — tool calling capabilities, image formats, streaming — and implementations can drift. Real adoption, measured by the number of gateways and platforms exposing it, will be the true test. For now, the two production implementations (Vercel, Databricks) are an encouraging signal, but the road to an effective standard runs through the major providers that are still missing.
The timeline also matters. Open Responses was first announced in January 2026; the donation eight months later is the moment it stops being an OpenAI project and starts being a community one. For developers, that is the signal to treat the spec as durable rather than experimental — but only where a real implementation exists, not everywhere it is merely announced. The gap between a press release and a working endpoint remains the most expensive mistake an integration team can make.
Verdict
If you build applications that call multiple models, adopt Open Responses where it is supported — through a gateway like Vercel AI Gateway or a platform like Databricks Model Serving — to shrink vendor-specific integration surface, while keeping an in-house abstraction for the edge cases the spec does not cover. If you are locked to a single provider, the benefit is smaller: track the spec’s evolution, but do not migrate for the sake of the standard alone. If you are choosing an agent stack today, favor components built on AAIF-governed standards — MCP, A2A, and now Open Responses — because that is where portability is being built. The battle between OpenAI’s, Anthropic’s, and Google’s dialects is not over, but for the first time, the cost of fragmentation falls on someone other than you.