Google releases Gemini 3.1 Flash Image, claims Pro-level quality at $0.50 per 1M tokens

Google has released Gemini 3.1 Flash Image, internally codenamed "Nano Banana 2," an image generation and editing model with a 131,000 token context window. The model is priced at $0.50 per million input tokens and $3 per million output tokens.

According to Google, the model delivers "Pro-level visual quality at Flash speed," positioning it as a faster, more cost-efficient alternative to its premium image models. The company claims the model combines advanced contextual understanding with fast inference, making complex image generation and iterative edits more accessible.

Technical specifications

Gemini 3.1 Flash Image supports:

• Context window: 131,000 tokens
• Multimodal input/output (image generation and editing)
• Configurable aspect ratios via the image_config API parameter
• Released: June 18, 2026

The model is available through OpenRouter, which routes requests across multiple hosting providers based on performance and pricing optimization.

Pricing comparison

At $0.50 per 1M input tokens and $3 per 1M output tokens, Gemini 3.1 Flash Image is positioned in the mid-tier pricing range for image generation models. OpenRouter reports that effective pricing can be 60-80% lower when prompt caching is applied for repeated context.

The pricing structure suggests Google is targeting production workloads that require both quality and cost efficiency, particularly for applications involving iterative image editing where context reuse is common.

What this means

Gemini 3.1 Flash Image represents Google's push into faster, more affordable image generation without sacrificing quality claims. The 131K context window is notably large for an image model, potentially enabling more complex multi-turn editing workflows. However, Google has not released benchmark comparisons against competing image models like DALL-E 3, Midjourney, or Stable Diffusion variants, making it difficult to independently verify the "Pro-level quality" claim. The model's real-world performance and adoption will depend on how it stacks up in user testing against established alternatives.