Nano Banana 2, measured: 11 seconds, and the only model that honours camera angles
Nano Banana 2 (Gemini 3.1 Flash Image) was the fastest model we measured, and the only one that actually produced the camera angle we asked for. Its limits are equally clear.
Measured
| Item | Measured value |
|---|---|
| Latency | ~11s (10–13s) |
| Cost per image | $0.068 |
| Output | JPEG |
| Measured on | 2026-08-28 (via OpenRouter) |
That is less than half of GPT Image 2's ~24s. On iterative work the gap becomes working time. The per-image cost, however, is 3.6× higher.
Angles actually work
Asking for 30° or 60° produced those angles. On other models the same instruction drifts back to a frontal view. For building shots around a fixed camera angle, this is decisive.
Eleven seconds barely feels like waiting.
But it emits one image per run. Five candidates means five runs, and the gap narrows.
Layouts and text break down
Multi-panel layouts and embedded text fall apart here, where GPT Image 2 follows them cleanly. The strengths are inverted, so expect to use both across a single project.
Accepted parameters
| Parameter | Accepted values |
|---|---|
| Aspect ratio | 1:1 / 2:3 / 3:2 / 3:4 / 4:3 / 4:5 / 5:4 / 9:16 / 16:9 / 21:9 |
| Resolution | 512 / 1K / 2K / 4K |
| Images per run | 1 only |
| Reference images | up to 14 |
On the other hand, you can choose the resolution — 512 through 4K — where GPT Image 2 has no resolution parameter at all.
Use it when the angle matters. Use something else when the arrangement matters.
Good for
- Iterating quickly
- Shots built around a specific camera angle
- When you want to pick the resolution
Bad for
- Getting several candidates from one run
- Multi-panel layouts or embedded text
Frequently asked
- How fast is Nano Banana 2?
- About 11 seconds (10–13s) measured. Under the same conditions GPT Image 2 took ~24s and Seedream 5.0 Lite ~41s.
- Can it return several images per run?
- No — one per run. If you need candidates, run count translates directly into cost.
- Does it keep the same person?
- Conditionally. If you need reliable identity from a single reference, GPT Image 2 was more stable. The conditions are in our comparison article.
