Google released Nano Banana 2 Lite, a Gemini image model built for speed and volume, and opened Gemini Omni Flash, its video generation and editing model, to developers through the Gemini API. Google announced both models on its Google Developers Blog on June 30, 2026. The releases target a specific gap: teams running high-throughput creative pipelines who need cheaper, faster generation without switching providers.
Nano Banana 2 Lite (model id gemini-3.1-flash-lite-image) produces text-to-image outputs in four seconds at $0.034 per 1,000-pixel image, according to Google’s own published benchmarks. The company positions it as a direct replacement for the original Nano Banana (gemini-2.5-flash-image), telling developers they can swap the model id and get faster, cheaper results immediately. That upgrade path matters more than the raw numbers: Google is trying to keep existing Nano Banana traffic in-house rather than lose it to a competing low-cost image API.
The model sits at the bottom of a four-tier Nano Banana lineup. Nano Banana 2 Lite optimizes for near-real-time, high-volume workflows. Nano Banana 2 (gemini-3.1-flash-image) is the general-purpose option balancing quality and latency. Nano Banana Pro (gemini-3-pro-image) handles complex, professional work where accuracy outweighs speed. The original Nano Banana is now labeled legacy, with Google recommending every remaining user migrate to the Lite tier.
Google’s benchmark chart compares Nano Banana 2 and 2 Lite against unnamed competitor image models on Elo score, latency, and cost per image. The blog post does not name which competitor models were tested, disclose sample sizes, or link to a reproducible benchmark suite. Readers evaluating the cost claim have Google’s number and no independent verification.
Gemini Omni Flash (gemini-omni-flash-preview) is the second release. It moves from the Gemini app and Google Flow, where Google introduced it at I/O, into the Gemini API and AI Studio as a public preview available to any developer. Omni Flash generates and edits video from a mix of text, image, and video inputs, and Google prices it at $0.10 per second of output, matching Veo 3.1 Fast.
Conversational editing is the feature Google is emphasizing: a developer can request a change to an existing video clip in natural language rather than re-prompting from scratch. Google’s Interactions API layers session memory on top, letting a user stack up to three sequential edits within one context. Google built three demo apps (Anywhere, Space Lift, and Omni product studio) specifically to show Nano Banana 2 Lite generating a still image and Omni Flash animating it, arguing the two models are meant to be chained rather than used in isolation.
The preview carries real limitations that Google discloses directly. Omni Flash currently caps generation at 10 seconds. Audio references and scene extension are not supported in the API. Video references under 3 seconds are accepted by the schema but not correctly processed, an unresolved bug Google is flagging rather than hiding. Character consistency across scene changes and camera pans is described as limited. None of these caveats are disqualifying for prototyping, but they rule out Omni Flash for finished long-form video work today.
Both models carry SynthID watermarking, and Google points to verification through the Gemini app, Gemini in Chrome, and Search as the provenance layer. That places the burden of proving AI origin on Google’s own detection tools rather than an open standard, a distinction worth noting given how often watermark verification depends on the platform doing the labeling.
For developers building generative media products, the practical decision is not whether to try these models but which tier fits which stage of the pipeline: Lite for drafting at scale, the standard tier for shipped output, Pro for anything requiring precision. Teams already paying for Veo at $0.10 per second should benchmark Omni Flash’s 10-second cap and editing limits against their actual use case before reallocating budget, since matched pricing does not guarantee matched capability.
Reported by Google on June 30, 2026.
