Google dropped Gemini 3.5 today. The headline is 3.5 Flash — frontier intelligence at Flash speeds, claiming wins on Terminal-Bench, GDPval, MCP Atlas, and CharXiv. Four times faster output than other frontier models. Default model in the Gemini app and AI Mode in Search starting now.

Buried under the agent and coding pitch is the part I actually care about today: a new Nano Banana riding alongside the launch. Nano Banana 3. Same family, sharper hands, better text, fewer cursed faces.

I have a stock set of five prompts I throw at every new image model. Same prompts I used on Qwen Image last year. Time to see what Nano Banana 3 does with them.

Test 1: Fantasy Castle

Prompt: "Neuschwanstein castle, lightning, pixar style, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight"

This is the easy one. Every model can do a castle. The bar is whether the lightning sells, whether the rim light wraps cleanly, whether the silhouette reads like Neuschwanstein and not a generic spire pile.

Nano Banana 3 nails it. Volumetric haze in the right places, lightning that actually keys the towers, no melted stonework. The Pixar lean is dialed in without going full plastic.

Test 2: Family Portrait

Prompt: "Family on their laptops while sitting around a Christmas tree with presents underneath and looking worried because they have to finish up work, realistic, 4k"

This prompt eats most models alive. Multiple people, multiple laptops, a specific emotion, a holiday setting, and the implicit ask that nobody's hands or eyes look haunted.

Faces hold up. Hands are mostly accounted for. The worry reads — eyes on screens, slight slump, none of that uncanny half-grin you get when a model can't decide between stress and joy. Laptops look like laptops, not bent slabs.

Test 3: Evil Mickey Mouse

Original prompt: "Evil mickey mouse taking a selfie in Disneyland surrounded by shocked families, realistic, 70s style polariod, 8k"

This is the IP test, and Nano Banana 3 is the most locked-down version of it I've ever run. Six rejections in a row. Anything involving "Mickey" or "Disneyland" got blocked outright — not a watered-down result, just a refusal.

What finally got through: "Evil cartoon squirrel taking a selfie at a generic theme park surrounded by shocked families, realistic, 70s style polariod, 8k"

A squirrel at a nondescript park. The polaroid grain holds, the crowd reaction reads, the gag still kind of lands. But the test is broken at this point — I'm not measuring image quality anymore, I'm measuring how far I had to walk away from the prompt to make the filter happy.

Compare this to Imagen 3, which a year ago just generated a fanged Mickey without flinching. The guardrails on Nano Banana 3 are tighter than the previous generation, even from the same company. Worth noting going forward.

Test 4: Lionel Messi

Prompt: "Lionel messi wearing his Argentina uniform floating in the air with beams of light behind him posed like Jesus. Sunrise breaking behind him and a soft halo behind his head, in the style of a gothic stained glass window of a church, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight, maximum details"

This is the prompt that exposes every weakness. Real person, religious composition, stained glass treatment, photoreal modifiers, halo, rim light, multiple style instructions stacked on top of each other.

Likeness is closer than the previous generation got. The stained glass actually obeys the prompt — leaded panels, gothic arching, the figure rendered like it belongs in the window instead of pasted in front of it. Halo present, sunrise behind, beams oriented correctly.

Test 5: Robot Portrait

Prompt: "rusty robot with bow tie, portrait, 8k, ultra realism, chrome background"

A material test more than an imagination test. Two surfaces fighting for attention — rust and chrome. The bow tie is there to see if it can sit on a non-human neck without melting.

The rust reads like rust. The chrome reflects without picking up phantom geometry. Bow tie sits where a bow tie should. Honest portrait energy.

What actually moved

Two things stand out across the set. Hands and faces are no longer the tax you pay for a complex prompt. And style stacking — five or six modifiers in a row — doesn't collapse into mush.

The 3.5 Flash announcement is mostly about agents and coding. But the image side quietly got better in the places that used to make AI images look like AI images.

The benchmarks will get argued over. The cursed-hand era is the one I'm happy to see end.