Google makes Interactions the default Gemini API

Google promoted its Interactions API to general availability on 22 June 2026, six months after it first appeared in beta in December 2025. For anyone building with Gemini, the practical line is short: this is now the recommended interface. The older generateContent endpoint still works, but the default has moved, and from now on new agent features ship only through Interactions. (The Decoder)

Logan Kilpatrick, Google’s developer relations lead, framed the moment as a step change for the platform: Interactions sets the stage for the new era of Agents.

What the interface actually changes

The shift is not just a rebrand of the same API. The shape of a request changes underneath. Where generateContent sent a list of free-form turns tagged as user or model, Interactions replaces that with typed steps — each action the model takes is its own defined step in the request, rather than text mixed into a single message blob. That lets developer tooling route, retry and stream individual actions instead of treating every turn as opaque text.

Google has updated AI Studio, the Gemini API documentation and the migration guide to be written against Interactions, so anything copied from the docs today will be the new shape. The older generateContent endpoint is not being switched off, but it has stopped being where new work happens — a maintenance surface, not a development one.

Capabilities that signal where Google is heading

Interactions ships with the agent features Google has been assembling over the past year, all now first-class parts of the default API rather than separate products to bolt on afterwards.

• Managed agents with Linux sandboxes — long-running agents get an isolated environment to execute code against, rather than the developer provisioning their own.
• Background execution — long-running tasks can be kicked off and polled, rather than held open inside a single HTTP request.
• Tool chaining with Google Search and Maps — agents can compose multiple Google tools in a single step rather than wiring each one by hand.
• Media generation — images, music and speech are native outputs, not a separate model call.
• Two cost profiles — a Flex mode pitched at roughly half the price of standard, and a Priority mode optimised for speed.

The mix tells a story. Google is selling agent infrastructure now, not just models. A UK team picking Interactions over generateContent is not just changing request shape — they are picking up sandboxed execution, asynchronous work and built-in tool routing as part of the deal. Whether that bundle is worth it depends on the workload, but the direction of travel is no longer in doubt.

The shape of the API is following the shape of the work. An agent that calls tools, runs for hours, and produces files is not a chat completion, and the interface is finally being built to match.

What to watch

The Google announcement is part of a wider pattern. OpenAI’s Responses API, Anthropic’s tool platform and now Google’s Interactions are all converging on the same answer: stateful, agent-native interfaces where the request describes a sequence of typed actions rather than a chat log. Two things to keep an eye on over the next quarter.

• Schema convergence. The big three labs now all speak the same general shape — typed steps, managed execution, tool chaining. UK teams building across providers should expect a portability layer to emerge, either as a community library or as a standards-body effort. The first credible one would be the more interesting story.
• What generateContent actually receives. Google has framed the old endpoint as still working, but in practice the surface area of new features is now Interactions-only. Watch for the first moment a flagship capability lands without a generateContent equivalent — that is the day the old API becomes legacy in fact as well as in name.

For teams shipping production agents on Gemini today, the migration is straightforward and Google has done the documentation work. For everyone else, the headline is simpler: the major labs have agreed, quietly, on what an agent API should look like, and the conversation has moved on from whether to build one to how they will converge.