Google had created a splash in the image generation space with Nano Banana, and it’s hoping to do an encore with Nano Banana 2.
The company’s newest model, Nano Banana 2 — formally known as Gemini 3.1 Flash Image Preview — has debuted at the top of Artificial Analysis’ Text to Image Leaderboard, edging out OpenAI’s GPT Image 1.5 (high) with an ELO score of 1,272 versus 1,268. The result is a meaningful milestone: Google’s Flash-tier model has beaten out the competition from OpenAI while also surpassing its own predecessor, Nano Banana Pro (Gemini 3 Pro Image), which ranks third on the same leaderboard with an ELO of 1,220.
Strong Performance Across Both Generation and Editing
The leaderboard results tell a nuanced story. On the Text to Image side, Nano Banana 2 is the clear leader. On the Image Editing Leaderboard, however, the picture shifts: GPT Image 1.5 holds first place with an ELO of 1,268, Nano Banana Pro comes in second at 1,250, and Nano Banana 2 secures a respectable third at 1,228. Still, landing in the top three on both leaderboards — against a competitive field that includes Black Forest Labs’ FLUX.2 models, xAI’s grok-imagine-image, and ByteDance Seed’s Seedream series — is a strong showing for a Flash-tier model that Google is positioning as a speed-first offering.
Nano Banana 2 Pricing
Perhaps the most compelling part of the Nano Banana 2 story isn’t the performance — it’s the price. At $67 per 1,000 images via the Gemini API, Nano Banana 2 costs roughly half what developers pay for either Nano Banana Pro ($134/1k) or OpenAI’s GPT Image 1.5 ($133/1k). For image editing workflows, it also significantly undercuts Black Forest Labs’ FLUX.2 [max], which runs $140 per 1,000 images. That kind of pricing, paired with top-tier rankings, makes Nano Banana 2 one of the most attractive value propositions in frontier image generation today.
What’s New Under the Hood
According to Google DeepMind, Nano Banana 2 brings several meaningful upgrades over the original Nano Banana. The model is grounded in Gemini’s real-world knowledge base and can draw on live information and images from web search to more accurately render real-world subjects — a capability Google is highlighting with a demo called “Window Seat,” which generates photorealistic window views from any location in the world using live local weather data at up to 4K resolution.
Beyond world grounding, the model supports subject consistency across up to five characters and 14 objects in a single workflow, making it more viable for storyboarding and narrative production. It also delivers improved instruction following, production-ready specs ranging from 512px to 4K, and an overall visual fidelity upgrade with sharper details, richer textures, and more vibrant lighting.
On the text rendering front, the model can generate accurate, legible text within images — useful for marketing mockups or greeting cards — and supports in-image text translation for localizing content across languages.
Wide Rollout Across Google’s Ecosystem
Google is not treating this as a quiet research release. Nano Banana 2 is rolling out as the new default across the Gemini app, Google Search (including AI Mode and Lens, now available in 141 countries and eight additional languages), and Google’s video generation platform Flow. Developers can access the model in preview via Google AI Studio and the Gemini API, as well as through Vertex AI on Google Cloud.
CEO Sundar Pichai called it Google’s “best image model yet,” pointing to its world understanding and real-time grounding as the core differentiators. For existing Google AI Pro and Ultra subscribers, Nano Banana Pro remains accessible for specialized high-fidelity tasks via the Gemini app’s regeneration menu.
SynthId Watermarking
Google is also continuing to invest in content authentication alongside the launch. The company’s SynthID watermarking technology has now been used over 20 million times since November to help users identify AI-generated images, video, and audio. Going forward, Google plans to pair SynthID with C2PA Content Credentials in the Gemini app, providing users with richer context about not just whether AI was used, but how.
The Bottom Line
Nano Banana 2 arrives as a direct challenge to the assumption that faster, cheaper models have to compromise on quality. By ranking first in text-to-image generation and third in image editing — at half the price of its closest competitors — Google has made a credible case that its Flash-tier model belongs at the frontier. For developers and enterprises evaluating image generation APIs, the cost-performance ratio makes Nano Banana 2 difficult to overlook.