Google ships the Gemini Interactions API to GA: the server holds state and runs your agents

The old pattern of resending the entire conversation state on every AI call is changing. Google has released the Interactions API, a unified interface for Gemini models and agents, to general availability. The server holds conversation state directly, runs work asynchronously in the background, runs agents in remote Linux sandboxes, and cuts cost by 50% on the Flex tier. ASAP summarizes the announcement from the primary source.

State moves from the client to the server

The Interactions API is Google's unified interface for Gemini models and agents. The heart of the change is a shift in who owns state. The server holds conversation state directly, runs asynchronously with background=True, combines built-in tools with custom functions, and supports multimodal generation. It started as a public beta in December 2025 and has now moved to general availability.

Why the state shift matters

The old pattern shipped the whole accumulated conversation on every call. The longer a session ran, the heavier each request became, and the burden of managing state piled up on the app side. When the server holds state, that weight moves to the other side of the wire. Asynchronous execution with background=True is a further signal in the same direction. Instead of a synchronous call-and-wait model, the server runs work for long stretches while the app merely collects the result. In effect, the API moves from behaving like a function toward behaving like a process.

Managed Agents work in remote Linux sandboxes

Managed Agents provision remote Linux sandboxes where an agent can reason, execute code, browse the web, and manage files. Deep Research comes in two versions, one tuned for speed and one for depth, with collaborative planning and native charts and infographics. Built-in tools such as Google Search and Google Maps combine with custom functions, and tool results can return images alongside text.

Why the schema goes "From Roles to Steps"

Media generation includes the Nano Banana 2 image model, the Lyria 3 music model, and expressive multi-speaker speech. The detail worth pausing on is that the schema changes "From Roles to Steps," so each action becomes a typed step. Older conversational APIs modeled the world as a sequence of messages exchanged between "roles," namely the user and the model. A step-centric schema instead treats each action, such as running code, searching, or generating an image, as its own typed unit. That is a natural fit for an agent that walks through several tools in sequence. The API supports Python and JavaScript SDKs and integrates with LiteLLM, Eigent, and Agno.

Cost and retention: Flex cuts 50%, paid tier keeps 55 days, legacy stays

The Flex tier reduces cost by 50%, and the paid tier retains data for 55 days. The Interactions API becomes the default across Google AI Studio and all official documentation. The legacy generateContent API remains fully supported, Google says.

What builders should weigh before adopting it

Server-held state with 55-day retention is a double-edged sword for any service. The app no longer has to manage conversation state itself, so the implementation gets lighter. But if conversations that carry personal data live on a vendor's servers for 55 days, privacy policies and cross-border transfer disclosures need a second look. On cost, the 50% Flex discount is attractive for early-stage experimentation, yet because it is a latency-tolerant tier, any service where real-time response matters has to weigh the trade-off against the paid tier. Since legacy generateContent stays fully supported, there is no rush to migrate wholesale; attaching the new agent features selectively, starting where they are actually needed, is the pragmatic path.

Limitations and open questions

What this announcement makes clear is the direction, not the stability of the fine print. Operational conditions such as the Flex tier's real-world latency, sandbox concurrency limits, and whether the 55-day retained data feeds training are hard to judge from the announcement alone. The more state the server holds, the greater the vendor lock-in. Once state and execution are tied to Google's infrastructure, the cost of switching to another model rises above what it used to be. Convenience and lock-in usually arrive together.

Source: ASAP summary of Google's Gemini Interactions API general availability announcement (public beta December 2025, now GA; server-held state and background=True async execution, Managed Agents in remote Linux sandboxes, built-in Google Search and Maps combined with custom functions, Deep Research in speed and depth versions, Nano Banana 2, Lyria 3 and multi-speaker TTS, Flex tier 50% cost reduction, paid-tier 55-day retention, Python and JavaScript SDKs, legacy generateContent fully supported).