Image Generation Model Comparison on One Prompt

What OpenRouter's image generation benchmarks compare

OpenRouter's Image benchmarks page sends the same prompt to the image generation models it lists and shows the results side by side. The page sits under "Media benchmarks", with the line "Compare generation quality across media models". For a developer or product owner who needs to choose an image model, it puts a quality signal, a price and a time into one row.

The listing sorts by Correctness, Cost or Generation time, and an Author filter narrows it by author. The tabs at the top read Image, Video, Speech and Memes, and all the figures quoted are from the Image tab. The saved text does not show what the other three hold.

The capture used here is the Full Wine Glass page, saved on October 14, 2026. It lists 44 rows, 42 of them with a score out of 4, and two names appear twice, so the count is 44 listings, not 44 different models. The page changes over time, so every figure below belongs to that capture.

Test groups run from wine glasses to image editing

The page sorts its tests into seven groups, from Improbable scenes to Consistency. The table lists each group with the tests named under it.

Test groups and the tests named under each

Group	Tests
Improbable scenes	Full wine glass / Closed umbrellas
Count	Three fingers / Cards & dice
Text	Long exact string / Multiple languages
Spatial relations	Occlusion / Mirror
Negation	Plain zebra / Times Square
Editing	Minimal diff / Object removal / Person removal
Consistency	Product placement / Four references

Negation is a group of its own, separate from Spatial relations. Every score, price and time quoted comes from the Full wine glass test, so none of them describes how the models do on counting, text, editing or consistency.

How the Full Wine Glass test checks a model

The Full Wine Glass prompt asks for white wine filled to the rim of an upright glass on a plain wooden table. The page adds that the shot is straight-on at eye level, the glass is not tilted, and nothing else is in the frame.

OpenRouter gives a reason for this prompt. Almost every wine photo in training data shows a glass about a third full, so models default to a third-full glass. A glass filled to the rim shows whether the model follows the prompt or the pattern it has seen most often.

Each row shows a score out of 4, written n/4 beside the model name. The saved text does not say how the score is produced, how many images stand behind it or who judges.

Six models score 4/4 on the Full Wine Glass test

Six rows reach 4/4, 29 sit at 3/4, six at 2/4 and one at 1/4, and two rows have no score. The two unscored rows are FLUX.3 Image and Recraft V4.1 Flash. The table lists the six top rows with the price and time beside each.

Rows scoring 4/4 on the Full Wine Glass test (saved October 14, 2026)

Model	Score	Price	Generation time	Author
Nano Banana 2	4/4	$0.069	8.5s	Google AI Studio
GPT-5.4 Image 2	4/4	$0.006	20.5s	OpenAI
GPT-5 Image	4/4	$0.17	42.9s	OpenAI
GPT Image 2	4/4	$0.006	16.3s	OpenAI
Qwen Image 3	4/4	$0.030	59.6s	Alibaba
Riverflow V2.5 Pro	4/4	$0.25	152.0s	Sourceful

For a team that wants a full score at the lowest price, two rows cost $0.006: GPT-5.4 Image 2 and GPT Image 2. GPT Image 2 is the quicker of the two, at 16.3s against 20.5s.

For a team that cares more about waiting time, Nano Banana 2 is the fastest 4/4 row at 8.5s and $0.069. The name Nano Banana 2 appears twice in the listing; the 4/4 row is the one that takes 8.5s, and the other row scores 3/4 at 9.0s.

Riverflow V2.5 Pro scores the same 4/4 as the cheapest rows but costs $0.25 and takes 152.0s. Price and waiting time do not rise with the score.

Prices run $0.006 to $0.25, times 1.2s to 208.2s

Across the 44 rows, the price runs from $0.006 to $0.25 and the generation time from 1.2s to 208.2s, and score, price and time move independently of one another. The cheapest rows are the two $0.006 rows, and both score 4/4. The highest price, $0.25, appears on two rows: Riverflow V2.5 Pro at 4/4 and Recraft V4 Pro at 3/4.

The fastest row is Recraft V4.1 Flash at 1.2s and $0.007, but it has no score, so there is no quality result to weigh its speed against. The slowest is Riverflow V2 Pro at 208.2s and $0.15, which scores 3/4. At the same $0.25, Recraft V4 Pro takes 19.1s and Riverflow V2.5 Pro takes 152.0s.

The table below adds rows from the FLUX, Nano Banana, Recraft and Riverflow families. They are ordered from the quickest to the slowest.

Selected rows from the Full Wine Glass listing, ordered by generation time (saved October 14, 2026)

Model	Score	Price	Generation time
Recraft V4.1 Flash	none	$0.007	1.2s
Nano Banana 2 Lite	2/4	$0.034	2.7s
FLUX.2 Klein 4B	3/4	$0.014	3.1s
Recraft V3	2/4	$0.040	4.5s
Nano Banana	2/4	$0.039	5.8s
FLUX.2 Pro	3/4	$0.030	7.2s
Nano Banana 2	4/4	$0.069	8.5s
FLUX.2 Flex	3/4	$0.050	10.6s
FLUX.2 Max	3/4	$0.070	16.5s
Recraft V4 Pro	3/4	$0.25	19.1s
Riverflow V2.5 Pro	4/4	$0.25	152.0s
Riverflow V2 Pro	3/4	$0.15	208.2s

The FLUX.2 rows range in price from $0.014 to $0.070 and all score 3/4. Nano Banana 2 Lite and Nano Banana both score 2/4, and the full-score Nano Banana 2 costs more than either. These figures describe one prompt, so a different task can order the same rows differently.

One test does not name the best image generation model

The score covers one prompt, so it cannot name an overall best image generation model. The saved page ranks the rows within the Full Wine Glass test only. It carries no figures for the other tests and shows no overall ranking across tests, and the saved text does not say whether one exists elsewhere on the site.

The single prompt is not the only limit. Three more come from what the saved text leaves out or from what changes over time:

• The saved text does not say how the n/4 score is produced, so a 4/4 means "top score on this prompt", not a verified quality grade.
• The saved text shows no free-tier marker, and every row has a dollar price.
• Prices and times change, so each figure here belongs to the October 14, 2026 capture and should be checked on the page before it informs a purchase.

For a developer or product owner, one sensible use of the page is to shortlist: keep the rows that reach 4/4 here, then run them on a prompt close to the real job. The page also lists text, editing and consistency tests, which are worth checking when the real job involves those tasks.

When sakutto.ai fetched the same page again on November 6, 2026, the Full Wine Glass list had grown to 55 rows and 9 of them scored 4/4. Here is how the counts compare with the October 14 capture.

Full Wine Glass rows by score (October 14 vs November 6, counted by sakutto.ai)

Score	October 14, 2026	November 6, 2026
4/4	6	9
3/4	29	37
2/4	6	6
1/4	1	1
No score shown	2	2
Total	44	55

The three new 4/4 rows are Nano Banana 2.1 ($0.035, 12.8s), GPT Image 2.5 Flare ($0.006, 9.4s) and GPT Image 2.5 Sunburst ($0.006, 12.3s). Models scoring 4/4 at $0.006 went from two to four, and the fastest of them is GPT Image 2.5 Flare at 9.4 seconds. Recraft V4.1 Flash, which showed no score on October 14, now shows 3/4 at 1.2 seconds.