Google ships Nano Banana 2 Lite and Gemini Omni Flash to developers, with a 4-second image path into video

The latency numbers are a frontend budget, not a benchmark

The headline spec for Nano Banana 2 Lite is a 4-second text-to-image output at $0.034 per 1K-resolution image. Google frames this as "rapid ideation and high-velocity developer pipelines where speed and cost are the primary constraints."

For anyone building an interface around generation, 4 seconds is the number that decides whether you show a spinner or a live preview. It sits in the awkward middle: too long to feel instant, short enough to justify an interactive drafting UI where a user tweaks a prompt and re-runs repeatedly. The cost figure matters for the same reason — at 3.4 cents an image, a UI that generates on every keystroke pause or slider change becomes financially defensible in a way a premium model would not.

Google explicitly positions this as the drop-in replacement for the original Nano Banana (gemini-2.5-flash-image), telling developers they can "swap it out now." That's a migration message aimed at existing frontend integrations, not just new builds.

The Interactions API and the three-edit ceiling

The most concrete frontend detail is quiet: the Interactions API maintains session history and context so "users can stack up to three sequential edits." That is a real constraint to design around, not a marketing point.

Conversational editing — the premise of Gemini Omni Flash — only works if the UI carries state across turns. A three-edit stack means the interface needs to communicate that budget to the user: what edits have been applied, when they're approaching the limit, and how to branch or restart. Building a multi-turn editor without surfacing this ceiling would produce confusing dead ends.

Omni Flash generates 10-second videos at $0.10 per second, which Google notes is "the same as Veo 3.1 Fast." Pricing parity with an existing model gives frontend teams a known cost baseline to reason about per-render.

The chaining pattern is the actual product Google is selling

The real magic happens when you chain these models together. Use Nano Banana 2 Lite as a high-speed image generation model, then pass that image as a reference to Gemini Omni Flash to animate it into a high-quality video.Montana Labs

The three demo apps all encode the same UI flow: generate a still with the fast image model, then let a click or button promote that still into an animated clip. Anywhere transports a selfie to landmarks then animates on click; Space Lift generates room redesigns then plays a cinematic showcase on tap; Omni product studio turns static images into e-commerce videos.

Each is a two-stage interaction where the cheap, fast model handles the browsing phase and the expensive video model is invoked only on explicit user commitment. That's a sensible pattern for controlling cost in a media UI — you don't animate everything, you animate the one thing the user picked.

The limitations that shape what you can ship today

Omni Flash is in public preview, and the caveats directly constrain UI scope. Videos are capped at 10 seconds. Audio reference uploads and scene extension aren't supported in the Gemini API yet. Video references up to 3 seconds are accepted by the API schema but "are not correctly processed by the model at this time" — a trap for anyone who validates against the schema and assumes it works.

Google also flags that character consistency degrades across scene changes and panning. For a frontend that promises a coherent character across multiple generated clips, that limitation defines the boundary between a demo and a shippable feature.

Both models carry SynthID watermarking, verifiable through the Gemini app, Chrome, or Search — a transparency mechanism that lives outside the app you build but is worth surfacing to users generating shareable media.

What the two-model split asks of media UI teams

The specific implication of this release is that Google is handing frontend teams a tiered generation stack and expecting the interface to route between tiers. Nano Banana 2 Lite for volume drafting, Nano Banana 2 as the generalist, Nano Banana Pro for accuracy-critical work, and Omni Flash for the video payoff.

Building well against this means the UI, not the model, decides when speed matters and when quality matters — cheap and fast for the exploratory phase, expensive and slow at the moment of user intent. The demo apps are Google's way of showing that the product is the orchestration between these models, and the frontend is where that orchestration becomes visible to the user.