Google's newest image model costs about half of what the last one did, and it drops more than just the per-image price. Anyone still calling Nano Banana 2 through the API has 23 days to move.
Nano Banana 2.1 Is Out and Half the Price
Nano Banana 2.1 reached general availability on October 6 with the model id gemini-nano-banana-2.1, an incremental name for what Google describes as a full step up in quality. The pitch is simple: better images, lower sticker price. A 1K image now runs $0.0336, down from $0.067 under Nano Banana 2. The 2K tier sits at $0.0504, and 4K lands at $0.0756, which is roughly half the earlier $0.151.
Billing is where the math gets less tidy. Image output got cheaper, but input pricing went up three times, now at $1.50 per million tokens, and the model thinks more by default. So if your workflow leans on heavy prompts or long conversations with the model, the savings on the picture itself can shrink. Always check the full bill, never the headline number.
Google says the upgrade covers visual quality, how closely the model follows your prompt, multi-turn character consistency, and text rendering. That last point matters for anyone who's tried to put a label or a sign into a generated image and watched the letters melt. Better text rendering turns image generation from a toy into something you can use on a poster.
Who Has 23 Days to Migrate Off Nano Banana 2
Here's the part with real consequences for developers. Google deprecated Nano Banana 2, which carries the model id gemini-3.1-flash-image, on the same day it shipped 2.1. The older model goes dark on October 29, 2026, a 23-day window from launch.
If you have production code pointed at the old model id, that's your deadline. Google's deprecation table names Nano Banana 2.1 as the recommended replacement, and every retired image id now points there. The migration is mostly a string swap, but you'll want to re-test your prompts, since a model change can shift how your favorite instruction lands. Skip it and your image calls start failing the morning of October 29.
What Set Benchmark Results Actually Hold
A bigger jump on the spec sheet is the reference-image handling. Nano Banana 2.1 can take up to 14 reference images in one request and keep up to 4 characters and 10 objects consistent across the whole set. For a comics workflow or a product campaign where the same mascot has to look identical in every frame, that consistency is the feature you'd pay for.
The model also connects to Google Search, so it can pull grounded detail into a prompt rather than invent it. Ask for a diagram of a specific device and it can look up the parts instead of hallucinating them. There's a thinking-effort dial too, with minimal, medium, and high settings, which lets you trade speed and cost against quality per job. Low stakes get the minimal setting. A hero image gets the high one.
The model shows up across the Google surface: the Gemini app, Search's AI Mode, AI Studio, Flow, and Stitch. That spread matters because it means the same engine behind a quick phone edit is the one filling in a design tool.
What Nano Banana 2.1 Costs in Practice
Cheaper per image sounds like a win, and for many users it's a win. A team generating a few hundred draft thumbnails a day will see the bill fall almost in half on output alone. That's the version of the story Google leads with, and it holds up if your prompts stay short.
The counterweight is the input side. Why does that matter? Tripling the input rate and defaulting to more thinking steps means image generation is no longer the whole cost of a request. If you feed in a wall of reference images and a long back-and-forth, the input tokens can carry a bigger share than the pictures. It pays to watch your token mix before you assume the new model is half price for your case.
How to Switch to Nano Banana 2.1
The new model is live now on the Gemini API and across the Gemini app, Search AI Mode, AI Studio, Flow, and Stitch. Update your model id from gemini-3.1-flash-image to gemini-nano-banana-2.1, then run your own prompts through both for a day before you cut over. Do that before October 29. After that date the old model stops answering, and any code still pointing at it simply breaks.
