GPT Image 2
GPT Image 2 is first on both Arena image leaderboards today: 1381 for text to image and 1463 for image editing, in both cases by a margin these two tables rarely show. One 1024x1024 image at medium quality costs $0.053, which is less than half of Nano Banana Pro.
- first on both Arena boards
- $0.053 per medium image
- 3840px ceiling on the long edge
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Forty five points of margin, which is a lot in these tables
On the text-to-image board the second row is 1336 and this model is 1381. On the editing board the second is 1439 and this is 1463. In a leaderboard of 77 models where most of the field sits inside fifty points, a forty five point gap means people are actually picking this output, not that an automated test scored it higher.
Read the name of that row, though: gpt-image-2 (medium). That is not a model, it is a serving configuration. OpenAI sells three quality tiers and its documentation prices all three at the same 1024x1024 size: low $0.006, medium $0.053, high $0.211. What came first on both boards is the middle one. The high tier appears on no leaderboard at all, so we do not know it is better and we do not write that it is.
Sizes, the one place in this table where exact numbers are published
The OpenAI guide gives three numeric constraints and all three are checkable: the long edge is at most 3840 pixels, both edges are multiples of 16, the long to short ratio is at most 3 to 1, and the total pixel count sits between 655,360 and 8,294,400. This is the only vendor in our table that gives pixel counts instead of saying 4K. Google writes 1K, 2K and 4K for every one of its models and nowhere says what 4K means in pixels. That is why the category table has no resolution column at all.
The column this model loses, and it is not our doing
Reference inputs. For anyone who needs one character or one product to hold across several images, that is the number that matters: how many reference images a single request accepts. Google publishes it for all three of its models and Black Forest Labs publishes it too. OpenAI does not. Its documentation shows a code sample with four input images, but nowhere says four is a ceiling, and we do not turn a sample into a limit. So the cell stays empty, ten points go, and this page says why.
Text inside the image: the only honest sentence anyone wrote
Every vendor writes a marketing adjective about text rendering. OpenAI wrote a limitation instead: the model "can still struggle with precise text placement and clarity". That is the only usable statement we read from any vendor in this area, which is why it is on the page. It is not a number, but it beats "advanced text rendering".
From Iran
We read the supported-countries list for the API today and Iran is not on it. That list lives on developers.openai.com, which opens, while openai.com itself returns 403 to our server. The distinction matters: the document speaks about the API, so whether the ChatGPT app opens from Iran is still something we do not know and do not write.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- blocked
- Payment
- no working route
- Free tier
- no
- Measured on
How we checked: We read the supported-countries list for the OpenAI API on developers.openai.com today and Iran is not on it. openai.com itself returns 403 to this server, so that list is the only readable document, and it speaks about the API rather than the ChatGPT app. No network test was run from inside Iran.
What it is good at
- First on both Arena image boards, by more than the gap between second and fifth place
- Medium quality costs less than half of Nano Banana Pro
- The only model here that publishes its allowed dimensions as pixel counts rather than labels
- Regional editing on an existing image, and a batch route at half price
Where it falls short
- No published ceiling for reference inputs, so that cell is empty in our table.
- The first place belongs to the medium tier; the high tier has no leaderboard row anywhere.
- High quality costs four times medium at the same size: $0.211 against $0.053.
- No weights are published, so there is no local route at all.
- Iran is not on the supported-countries list for the API and there is no official route.
- OpenAI itself says the model can still struggle with precise text placement.
Our take
If output quality is the only thing that matters and your payment route is already solved, this is the right pick today, and the rival charging twice as much scores below it. If your work is several images holding one consistent character, run a small test before committing: this is the one model whose reference ceiling you cannot look up.
Questions people actually ask
What is the best AI image generator
On human preference, GPT Image 2 leads both Arena image boards today. If your payment route or your data privacy is the constraint, Nano Banana 2 is cheaper and easier to reach, and FLUX.2 klein runs on your own machine.
How much does one image cost with GPT Image 2
At 1024x1024: $0.006 at low quality, $0.053 at medium, $0.211 at high. The batch route is half price. These numbers come from the cost table in OpenAI own documentation.
How many reference images can it take
OpenAI publishes no ceiling. Its documentation shows a sample with four input images, but nowhere calls four a limit, so we do not record it as one.
Can it be used from Iran
There is no official route. Iran is not on the supported-countries list for the OpenAI API, which we read ourselves today. We do not sell accounts and we do not suggest ways around it.
Sources
- OpenAI, gpt-image-2 model page vendor source 12 August 2026
- OpenAI, image generation guide, sizes and cost table vendor source 12 August 2026
- OpenAI, API pricing vendor source 12 August 2026
- OpenAI, supported countries and territories for the API vendor source 12 August 2026
- Arena, text to image leaderboard vendor source 12 August 2026