The best AI for generating images

Why GPT Image 2 leads despite an empty cell

Fifty of the hundred points in this architecture belong to the two Arena boards, and this model captures both. On text-to-image conversion it scores 1381.1, while the runner-up records 1263.8, and on editing it stands at 1462.7 against 1392.6. These are wide gaps, and neither confidence interval reaches the next model.

The notable part is that it won with an empty column. OpenAI does not publish a ceiling for reference-image capacity anywhere in its documentation; its sample code accepts four images, but a sample is not a specification ceiling, so that cell was left empty and this model gave up ten points. Even with that ten-point loss, it finishes eight points ahead of second place.

One operational condition is registered directly in the leaderboard row name and gets ignored in most sources: the row that took first place is "gpt-image-2 (medium)", the medium-quality configuration. The high-quality tier at the same size costs $0.211, four times as much, and has no leaderboard row at all. The figure recorded in our price column, $0.053, belongs to the configuration that actually earned the rank.

The lowest-cost route to a strong result

Nano Banana 2. Alongside Nano Banana Pro, it is the only candidate that fills all six columns of this architecture, and the only one that combines full coverage with a high score. It captures the reference column outright (ten object images) and undercuts the first-place model on price.

And here is the finding that changes a purchasing decision: Nano Banana Pro, Google expensive flagship model, ranks third in this table and sits below its own cheaper model. It is genuinely behind on text-to-image conversion, confirmed by the confidence intervals, level on editing, and costs twice as much. If you are choosing between the two Google models, the cheaper one is almost always the correct answer.

The bottom row is the only one you can download

FLUX.2 klein 4B ranks sixth, and that should not be read as weakness. It captures the price column outright at $0.014, roughly a quarter of the first-place cost, 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.

That means if your operational requirement is deploying the model on your own infrastructure, 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, meaning commercial use is permitted. Its quality trails the top of the table and that has to be accepted; in exchange it requires no account, no bank card, and no country of yours appearing on any list.

The Iran column creates no distinction today

This has to be stated plainly, because it cannot be read off the table itself. Three candidates carry data in this column: GPT Image 2 and both Google models. All three are in the identical situation: Iran is absent from their supported-country lists, and there is no official payment route. When every value in a column is equal, the scoring engine awards full marks to all of them. So this column produces no difference at all among the top three.

What this column actually does is dock ten points from the other three, not because their situation is worse but because their vendors have published no document to read at all. For FLUX we reviewed the full documentation index today: there is no terms page and no country list, and the terms URL on the main site returns a 404 status code. ByteDance has published nothing comparable for Seedream. In practice this column is measuring whether a vendor publishes a country list, which is not the same question as whether the service works from Iran. Until more documentation surfaces, that is the state of it, and we are not concealing it.

Two criteria deliberately not built into this architecture

First, resolution ceiling. Three vendors publish three incompatible specifications, and the three do not reconcile. Google labels its models "up to 4K" and nowhere publishes pixel dimensions for any tier from 0.5K to 4K. OpenAI is the most precise: 3840 pixels on the long edge and a total pixel count between 655,360 and 8,294,400. Black Forest Labs states 4 megapixels. A label with no number, an exact figure, and a third unit — no column can be built from that combination. These figures were kept as facts on each model own page, and no column was built.

Second, text inside the image. For Persian speakers this may be the single most important operational question, and it is exactly the one no vendor publishes a number for. Google describes its text rendering as advanced. OpenAI itself writes that its model can still struggle with text rendering. Black Forest Labs calls flex specialised for typography. These are three adjectives, and none of them is a measurement. A column built from adjectives is not a measurable criterion. The key for this metric exists in our data structure and stays empty until somebody publishes a number for it.

Why Midjourney has no presence in this architecture

Because no data has been extracted from it at all. midjourney.com returns a 403 status code to our server: the home page, the updates page, and its documentation alike. One difference from OpenAI is worth recording: with openai.com, robots.txt at least responds and does not forbid crawling, whereas midjourney.com/robots.txt is itself a 403, meaning even the rule we would be expected to follow cannot be read.

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

A row that ranked on half the columns

Seedream 5.0 Pro finishes fourth while filling only two of the six columns in this architecture. It ranks 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, even though FLUX filled four columns. Read this carefully: Seedream position in this table means it performs well on the two things measured, not that it is superior to FLUX.

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