Nano Banana 2

For years the rule held: the best image generation cost the most. The flagship tier produced the leaderboard-topping pictures; the cheap, fast tier gave you a worse result for less money. On February 26 Google released Gemini 3.1 Flash Image — codenamed Nano Banana 2 — and it landed first on Artificial Analysis' blind text-to-image arena, where humans pick the better picture without knowing which model made it. It edged out OpenAI's contender and beat Google's own pricier Nano Banana Pro, the model it was supposed to sit beneath, while costing roughly half as much per thousand images.

The headline 4.5-cent price buys a 512-pixel thumbnail; a usable 4K image costs fifteen cents.

The inversion is the story: top-rank quality moved into the budget lane. What didn't move is the headline price floated around the launch. The cheapest tier does run about four and a half cents an image, but that buys a 512-pixel thumbnail; a true 4K image costs fifteen cents, more than three times as much. The cents-per-picture figure and the resolution that makes it useful for commercial work don't belong to the same image.

Most people never chose it. Google made the model the silent default across the Gemini app, Search, Lens, Ads and its Flow video tool in 141 countries, so hundreds of millions of image requests quietly switched engines overnight. That reach, not the benchmark, is the real weight here — and it points at the commodity end of stock photography, where the work is generic and the markup was the product.

The leaderboard crown, though, was rented. Within weeks OpenAI's next model retook the top spot, which is the tell about this whole race: the rank changes hands monthly, so the durable shift isn't who is briefly best but that best-in-class image quality no longer carries a premium price.