PARALLELbenchmark
image model test — 13 August 2026

nano-banana-2 stays closer to the original, for a fraction of the price.

We remade 10 slides from 2 templates with both models and compared every result to the original. Start with the headline, then look template by template, then open any single image.

images made
18 of 20
templates
2
total spend
62.9¢
time spent waiting
11m 56s
the short version
gpt-image-2
matches original
92%
cost per image
5.9¢
wait per image
64s
delivered
10 / 10

Returned an image every time. Averages 64s and 5.9¢ per render.

nano-banana-2our pick
matches original
98%
cost per image
0.4¢
wait per image
9s
delivered
8 / 10

Returned nothing on 2 of 10 slides. Averages 9s and 0.4¢ per render.

Matches the original
average across every image
92%
98%
6 points
nano-banana-2
Wait per image
from request to finished file
64s
9s
7.3× faster
nano-banana-2
Cost per image
what one render costs
5.9¢
0.4¢
13.8× cheaper
nano-banana-2
Came back with an image
how often it delivered
10/10
8/10
2 more
gpt-image-2
gpt-image-2nano-banana-2longer bar = more of that thing
template 31929 — 2 slides

Betr NFL Global Workflow (Noah) 2026-08-12 1

2 slides sent to both models. gpt-image-2 returned 2, nano-banana-2 returned 1.

closer to original
gpt-image-2 on 1 of 2
what it cost
10.1¢ vs 0.5¢
Slide 01even
original
97%37s
97%10s
Slide 02gpt-image-2 only
original
70%69s
came back empty
no image5s
template 31992 — 8 slides

Arya(18-25)(Relationship)(stages of relationships)(Maribela)(Arya) 1

8 slides sent to both models. gpt-image-2 returned 8, nano-banana-2 returned 7.

closer to original
nano-banana-2 on 6 of 8
what it cost
49.3¢ vs 3¢
Slide 01gpt-image-2 only
original
100%76s
came back empty
no image4s
Slide 02nano-banana-2
original
97%49s
100%8s
Slide 03nano-banana-2
original
94%118s
95%10s
Slide 04even
original
97%64s
97%9s
Slide 05nano-banana-2
original
97%54s
100%8s
Slide 06nano-banana-2
original
89%48s
98%7s
Slide 07nano-banana-2
original
88%65s
97%10s
Slide 08nano-banana-2
original
92%59s
97%8s

Match is how close a render sits to the original slide, scored automatically with a perceptual hash. Where a model returned nothing, its slot is left empty rather than filled with a substitute.