Nano Banana 2 generates images at 1024×1024 resolution (1K) with a median end-to-end API latency of approximately two to four seconds under normal load conditions. This figure covers the full round-trip from request submission to first byte of the response, including model inference and network transfer of the base64-encoded image payload. The latency is not constant—it varies based on server load, prompt complexity, and whether the request includes reference images. Requests with four reference images and a complex prompt take longer than requests with a simple prompt and no references, because the model processes more input tokens before the generation pass begins.
At the 95th percentile, latency can rise to six to eight seconds during peak usage periods. If your application has strict latency requirements—for example, a user-facing feature where the image must appear within three seconds of the user's action—you should benchmark against the 95th percentile rather than the median to ensure the experience holds up under real traffic conditions. Building in a timeout and a graceful degradation path (such as showing a placeholder while the image loads asynchronously) is a standard practice for production integrations with generative APIs where tail latency can exceed the median by a meaningful factor.
For batch generation workflows that do not have real-time latency requirements, throughput matters more than per-request latency. Nano Banana 2's throughput is constrained by your project's rate limit, which is measured in requests per minute. Saturating the rate limit with concurrent requests allows you to maximize the number of images generated per unit of time, and the total generation time for a batch scales roughly linearly with the number of images divided by the rate limit. If you need higher throughput than a single project's rate limit allows, the API documentation describes the process for requesting a quota increase.