Firebase AI Logic, the SDK that lets you call Gemini directly from a mobile or web client without standing up a backend, is moving into General Availability at I/O 2026. It confirms an architectural choice Google has defended since the product was still called Vertex AI in Firebase: the model call lives inside the app, not behind a proxy, but under the control of security rules, usage monitoring, and automatic API key management.
Grounding with Google Maps: the day's highlight
The feature that deserves the most attention is Grounding with Google Maps, also GA inside the SDK. The principle is simple: before answering, the model consults Google Maps Platform and anchors its responses to real-world geospatial data. For anyone building a travel guide, a restaurant recommendation engine, or a local logistics app, it means text generation stops inventing addresses, opening hours, or names of nonexistent venues. You enable it by configuring the generative model (for instance Gemini 3.5 Flash) and, if useful, passing the user's latitude and longitude in the retrieval config, so the results pick up a geographic bias.
Grounding with Maps is available from the Firebase AI Logic SDKs for Android, iOS, Flutter, JavaScript and Unity. Google announced the same capability in parallel as part of the new agentic wave on Google Maps Platform, a signal that geolocation becomes a first-class asset for agents.
Nano Banana under programmatic control
The integration with Nano Banana, Google's in-house image editing model, gets an API that lets you control aspect ratio and size of generated images programmatically. Developers had been asking for this for months: the model produces high-quality images, but they needed a way to request 1080x1350 for an Instagram post or 1200x630 for an Open Graph card without resorting to textual workarounds in the prompt. Now the parameter is explicit.
Robustness for real apps
Two features that weigh less on the keynote but a lot on perceived quality: session resumption and context compression. If the network drops in the middle of a long conversation with the model, the SDK rebuilds state instead of starting over, and meanwhile reduces the history payload. These are the primitives that were missing to take conversational experiences into production on mobile devices on the move, where connection quality is the most unpredictable variable.
Security: template-only and authentication-mode
The template-only mode forces the client to execute only prompts stored server-side: SDK users cannot inject arbitrary prompts, eliminating an entire class of client-side prompt injection attacks. Authentication-mode, coming in the next few weeks, requires Gemini calls to include a valid Firebase Authentication token. Together they shift the security perimeter from the model to the contract.
Workspace as a conversational data source
The last piece, already mentioned in Firebase's I/O materials, is the connector to Google Workspace: apps can now query Gmail, Docs and Sheets in natural language, using Sign in with Google as the authorization layer. For many B2B applications this removes the need to build custom integrations: the conversational backend is already there, you just declare the permissions.
Why it matters
Firebase AI Logic moves from an experimental SDK to a production platform: GA of the Gemini 3.x models, Grounding with Maps in GA, explicit security, network resilience, Workspace integration. For anyone building vertical apps — travel, food, mobility, local retail — the cost of building an assistant that does not hallucinate drops dramatically. And for Google, it is the missing piece to push agents into the last mile of the client, where the real friction with the user lives.