GPT Image 2
GPT Image 2 holds first position on both Arena image leaderboards: a score of 1381 for text-to-image and 1463 for image editing, in both cases by a margin these two tables rarely record. One 1024x1024 image at medium quality is priced at $0.053, less than half of Nano Banana Pro.
- First position on both Arena leaderboards
- $0.053 rate 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. Change log
A forty-five-point margin, a substantial figure in these tables
On the text-to-image leaderboard the second-place row scores 1336, and this model scores 1381. On the editing leaderboard second place scores 1439, and this scores 1463. In a leaderboard spanning 77 models where most of the field clusters within fifty points, a forty-five-point gap indicates users are genuinely selecting this output, not that an automated test assigned it a higher figure.
The entry id, however, needs a precise read: gpt-image-2 (medium). This is not a standalone model, it is a serving configuration. OpenAI ships three quality tiers, and its technical documentation prices all three at the same 1024x1024 size: low at $0.006, medium at $0.053, high at $0.211. The configuration that took first position on both leaderboards is the middle tier. The high tier does not appear on any leaderboard at all, so no claim of superiority can be made and none is recorded here.
Dimensions, the one section of this table with exact published figures
OpenAI documentation records three numeric constraints, and all three are independently verifiable: the long edge caps at 3840 pixels, both edges must be multiples of 16, the long-to-short ratio caps at 3 to 1, and total pixel count falls between 655,360 and 8,294,400. This is the only vendor in this reference table that records pixel counts rather than a label like 4K. Google uses the labels 1K, 2K and 4K across all of its models and nowhere specifies the pixel equivalent of 4K. That is precisely why the resolution column is not defined in this category table at all.
The column where this model loses points
Reference inputs. For a workload that must hold one character or one product consistent across several images, this is the figure that matters: how many reference images a single request accepts. Google publishes this figure for all three of its models, and Black Forest Labs publishes it as well. OpenAI does not. Its documentation code sample shows four input images but nowhere declares four as a ceiling, and a code sample cannot be recorded as a specification ceiling. So this cell stays empty, ten points are lost, and this document records the reason.
Text inside the image: the one transparent statement in this domain
Every vendor records a marketing adjective about text rendering. OpenAI recorded a limitation instead: the system can still struggle with precise text placement and clarity. This is the one citable statement recorded by any vendor in this domain, which is why it is recorded here. It is not a figure, but it is more precise than advanced text rendering.
Reachability from Iran
The supported-countries list for the API has been extracted, and Iran is not recorded on it. This list is hosted on the developers.openai.com infrastructure, which is reachable, while the openai.com infrastructure itself returns a 403 status to our measurement server. This distinction matters: this document speaks strictly about the API, so whether the ChatGPT app is reachable from Iran remains unrecorded, and no claim is made about it.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
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: The supported-countries list for the OpenAI API was extracted from the developers.openai.com infrastructure, and Iran is not recorded on it. The openai.com infrastructure itself returns a 403 status to this measurement server, so that list is the only accessible document, and it speaks strictly about the API rather than the ChatGPT app. No network test has been run from inside Iran.
Documented strengths
- First position on both Arena image leaderboards, by a margin larger than the gap between second and fifth place
- The medium quality rate is less than half of Nano Banana Pro
- The only model in this table that publishes its allowed dimensions as pixel counts rather than labels
- Regional editing on an existing image, and a batch route at half the rate
Known weaknesses and limits
- No published ceiling for reference-input capacity, so that cell is recorded empty in this reference table.
- First position belongs to the medium tier; the high tier holds no leaderboard entry anywhere.
- High quality is priced at four times medium at the same size: $0.211 against $0.053.
- No weights are published, so no local deployment path exists at all.
- Iran is not recorded on the supported-countries list for the API and no official route exists.
- OpenAI itself states the system can still struggle with precise text placement.
Technical verdict
If output quality is the sole decision criterion and the payment route is already resolved, this model is the correct choice today, and the rival charging twice the rate scores below it. If the workload involves producing several images holding one consistent character, run a small test before full commitment: this is the one model whose reference-input ceiling is not specified.
Frequently raised questions, with documented answers
Which model is the best AI image generator
By human preference measurement, GPT Image 2 holds first position on both Arena image leaderboards. If payment route or data privacy is the binding constraint, Nano Banana 2 carries a lower rate and simpler reachability, and FLUX.2 klein is deployable on the user own infrastructure.
What is the per-image rate for GPT Image 2
At 1024x1024: low quality at $0.006, medium at $0.053, high at $0.211. The batch route records half that rate. These figures are extracted from the cost table in OpenAI official documentation.
How many reference images can this model accept
OpenAI has not published a ceiling. Its documentation code sample shows four input images but nowhere declares four as a ceiling, so this figure is not recorded as a specification limit.
Is this model reachable from Iran
No official route exists. Iran is not recorded on the supported-countries list for the OpenAI API; this list was directly checked and extracted. This report offers no account resale path and suggests no circumvention route.