The best AI for generating images

Why GPT Image 2 came first, with an empty cell

Fifty of the hundred points in this table belong to the two Arena boards, and this model wins both. On text-to-image it scores 1381.1 while the runner up sits at 1263.8, and on editing 1462.7 against 1392.6. Those are wide gaps and neither confidence interval comes near the next model.

The interesting part is that it won with an empty column. OpenAI does not publish a ceiling for reference images anywhere in its documentation; the sample code passes four images, but a sample is not a limit, so we left that cell empty and the model gave up ten points. Even ten points down, it finishes eight ahead of second place.

One condition is written into the leaderboard row itself and gets ignored almost everywhere: the row that came first is "gpt-image-2 (medium)", the medium quality tier. High quality at the same size costs $0.211, four times as much, and has no leaderboard row at all. The number in our price column is the $0.053 of the configuration that actually earned the rank.

The cheapest route to a good result

Nano Banana 2. Along with Nano Banana Pro it is one of only two candidates that fills all six columns, and it is the only one that combines full coverage with a high score. It wins the reference column outright with ten object images, and it undercuts the first-place model on price.

Then there is the finding that should change a purchase: Nano Banana Pro, Google expensive flagship, lands third here, below its own cheap model. On text-to-image it is genuinely behind and the confidence intervals confirm it, on editing the two are level, and it costs twice as much. If you are choosing between the two Google models, the cheaper one is almost always the answer.

The bottom row is the only one you can download

FLUX.2 klein 4B finishes sixth, and that should not be read as bad. It wins the price column outright at $0.014, roughly a quarter of what first place costs, and it is the only candidate that scores at all on weight openness. The other five all score zero there, because none of them has published weights.

So if your requirement is that the model runs on your own hardware, this table does not offer six options. It offers one. It runs on a graphics card with about 13GB of memory and its licence is Apache 2.0, so commercial work is allowed. Its quality sits well below the top and you have to accept that; in exchange it needs no account, no bank card, and no membership of anybody country list.

The Iran column separates nobody this month

This has to be said plainly because the table does not show it. Three candidates have data in this column: GPT Image 2 and both Google models. All three are in exactly the same situation, which is that Iran is absent from their supported-country lists and there is no official payment route. When every value in a column is equal, the engine awards all of them full marks. So this column creates no difference at all among the top three.

What it actually does is dock ten points from the other three, not because their situation is worse but because their vendors published no document to read. For FLUX we opened the full documentation index today: it contains no terms page and no country list, and the terms URL on the main site returns 404. ByteDance has published nothing comparable for Seedream. So in practice this column is measuring whether a vendor publishes a country list, which is not the same question as whether the thing works from Iran. Until more documents exist, that is what it is, and we are not hiding it.

Two criteria we deliberately did not build

First, resolution ceiling. Three vendors write three different things and the three do not add up. Google writes "up to 4K" for its models and nowhere prints pixel dimensions for any tier from 0.5K to 4K. OpenAI is the most precise: 3840 pixels on the long edge and a total between 655,360 and 8,294,400. Black Forest Labs says 4 megapixels. A label with no number, an exact number, and a third unit. You cannot build a column out of that, so we kept the values as facts on each model page and built no column.

Second, text inside the image. For Persian speakers this may be the single most important question, and it is exactly the one nobody publishes a number for. Google calls its text rendering advanced. OpenAI writes that its model can still struggle with text. Black Forest Labs calls flex specialised for typography. Those are three adjectives and not one measurement. A column built out of adjectives is not a column. The key exists in our data structure and stays empty until somebody prints a number.

Why Midjourney is not in the table at all

Because we have read nothing from it. midjourney.com returns 403 to our server: the home page, the updates page and its documentation alike. One difference from OpenAI is worth writing down: with openai.com at least robots.txt answers and does not forbid crawling, whereas midjourney.com/robots.txt is itself a 403, so even the rule we would be expected to follow cannot be read.

And unlike OpenAI, whose model holds first place on both boards, Midjourney appears on neither Arena image board: seventy seven models on text-to-image, fifty three on editing, and its name is on neither list. So we have no facts and no evidence, and we will not manufacture a row with no numbers in it. If you read somewhere that Midjourney is the best, ask where the number is.

One row that ranked on half a table

Seedream 5.0 Pro finishes fourth having filled only two of the six columns. It sits high on both boards, fifth on editing and eighth on text-to-image, and that alone was enough to edge past FLUX.2 max by four tenths of a point, while FLUX filled four columns. Read that carefully: Seedream position means it is good at the two things we measured, not that it is better than FLUX.

Those four empty columns are not our forgetfulness either. The Seedream 5.0 Pro page opens and is full of samples, but carries no model ID, no price, no reference ceiling and no date. The Seedream 5.0 page returns 200 with an empty body. This row coverage is exactly 50 out of 100, right on the boundary: one more empty cell and it would have dropped out of the table into the "not enough data" list beneath it.