Google's Gemini 3.1 Flash Image, which Google itself calls Nano Banana 2, generates and edits images at up to 4K and bills $67.00 per 1,000 images.
On the Artificial Analysis image arena it scores 1122 Elo, seventh overall.
The specification
Two latency figures appear above because they measure different things. 9.1 seconds is Artificial Analysis timing a generation. 17.7 seconds is the median end-to-end request on the served endpoint over a trailing thirty-minute window, which includes queueing. Neither is wrong; they answer different questions.
Where it sits
Nano Banana 2 is 66 Elo behind GPT Image 2.5 Flare and costs 68% less. It is 25 Elo behind Microsoft's MAI-Image-2.6 and costs 72% more.
That second comparison is the awkward one. On the arena's evidence, MAI-Image-2.6 is both better-rated and substantially cheaper. Nano Banana 2's case rests on what the arena does not score: 4K output, which MAI-Image-2.6 does not offer — it tops out at 1.5K — and image editing with a supplied reference.
Google also ships Nano Banana 2 Lite at 1091 Elo and $33.60 per 1,000, with a 3.1-second generation time. For thumbnail and preview work, the Lite tier is the better buy by a wide margin.
What it is for
If you need 4K, Nano Banana 2 is one of very few models that produces it, and it is the cheapest of those. If you are generating at 1K for a feed, you are paying a 4K premium you are not using — Muse Image at $10 per 1,000 or MAI-Image-2.6 at $38.90 will cost far less for output nobody is going to zoom into.
Elo and generation time are from the Artificial Analysis image arena, read 21 September 2026; price, context, output ceiling and request latency from the model's served endpoint on the same date. Resolution tiers are from Google's own model documentation. Every figure carries its source and date on the model page.
