Skip to content

Imagen 3.0 Generate

Overview

Imagen 3.0 Generate is Google’s dedicated high-fidelity image generation line rather than a general multimodal Gemini model. Google’s Imagen guide describes Imagen as a model family focused on realistic, high-quality image generation, while the deprecations page shows that the older Imagen 3 generation variants have already been superseded by newer releases. That makes this entry especially relevant as a legacy or migration document rather than a recommendation for fresh builds.

API base URL

https://generativelanguage.googleapis.com/v1beta

Example payload

{
  "model": "imagen-3.0-generate-001",
  "prompt": "A photorealistic portrait of a cyclist in rainy evening light",
  "config": {
    "numberOfImages": 1,
    "aspectRatio": "1:1",
    "imageSize": "1K",
    "outputMimeType": "image/png"
  }
}

Our Current Pricing

The exported model record includes nominal token-style pricing metadata, but the current MODEL_COSTS_OVERVIEW.md does not describe an exact active billing function for Imagen 3.0. Because the user asked for the repo-accurate picture, the most honest statement is that there is no finalized, separately documented pricing routine here that matches the level of detail available for GPT Image, Gemini image, or DALL·E.

The practical way to document billing for this model is to treat the record metadata as a provisional rate card:

output_tokens_per_image = 1290
output_cost_per_1m_tokens = 30 USD
estimated_cost_per_image = 1290 * 30 / 1_000_000 = 0.0387 USD

If you decide to bill from this metadata, the influencing parameters are mainly image count and any request options that change output size or model tier in your own product layer. Since the repo does not currently expose a richer formula here, the cleanest implementation is to persist the request, compute a deterministic estimate from the configured rate card, and tag the bill line as configuration-based rather than provider-metered.

If the provider later begins returning explicit usage, move billing to the usage object immediately and keep the record-level number only as a fallback.

Sources