Skip to main content
Workroom

GPT Image 2

Try itCompareRead the study

Rank

#1

Score

104

Price

$0.058

GPT Image 2 scores 104.4 on our photo-anchored scale, where a real photograph is fixed at 100. That sits above the scale's ceiling but is not a claim of parity: in blind tests on 48 commercial scenes its believability gap to the photographs, +0.09 with a 95% interval from −0.16 to +0.33, was not separated. It clearly beat Qwen Image 2.0, FLUX.2 Klein and Seedream 4.0.

Vendor
OpenAI
Family
GPT Image
Type
T2I
Modalities
→
License
Proprietary
Released
Apr 2026
Max resolution
Up to 4K
Edit modes
Inpaint · Reference
LoRA
No
Best for
Fashion and beauty, Interior, Street and architecture

Where it stands

A real photograph is pinned at 100, and every model sits by how often it won blind comparisons — the further left, the less often. Scroll sideways to see the rest of the ruler.

  • 117 · #1

    • Preferred over the real photograph in 1 of 3 blind comparisons

    • Preferred over the real photograph in 2 of 3 blind comparisons

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • Preferred over the real photograph in 2 of 3 blind comparisons

  • 114 · #2

    • The real photograph was preferred in 2 of 3 blind comparisons

    • Even in 3 blind comparisons

    • Preferred over the real photograph in 2 of 3 blind comparisons

    • Preferred over the real photograph in 3 of 3 blind comparisons

    • The real photograph was preferred in 1 of 3 blind comparisons

    • The real photograph was preferred in 3 of 3 blind comparisons

  • 114 · #1

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • Preferred over the real photograph in 3 of 3 blind comparisons

    • The real photograph was preferred in 1 of 3 blind comparisons

    • Even in 3 blind comparisons

  • 112 · #1

    • Preferred over the real photograph in 2 of 3 blind comparisons

    • Even in 3 blind comparisons

    • The real photograph was preferred in 2 of 3 blind comparisons

    • Even in 3 blind comparisons

    • Preferred over the real photograph in 2 of 3 blind comparisons

    • Preferred over the real photograph in 3 of 3 blind comparisons

  • 112 · #1

    • Preferred over the real photograph in 3 of 3 blind comparisons

    • Preferred over the real photograph in 3 of 3 blind comparisons

    • The real photograph was preferred in 1 of 3 blind comparisons

    • Preferred over the real photograph in 1 of 3 blind comparisons

    • Even in 3 blind comparisons

    • The real photograph was preferred in 3 of 3 blind comparisons

  • 91 · #1

    • Preferred over the real photograph in 1 of 3 blind comparisons

    • The real photograph was preferred in 2 of 3 blind comparisons

    • The real photograph was preferred in 1 of 3 blind comparisons

    • The real photograph was preferred in 3 of 3 blind comparisons

    • The real photograph was preferred in 2 of 3 blind comparisons

    • Preferred over the real photograph in 2 of 3 blind comparisons

  • 85 · #1

    • The real photograph was preferred in 3 of 3 blind comparisons

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • The real photograph was preferred in 3 of 3 blind comparisons

    • Preferred over the real photograph in 1 of 3 blind comparisons

  • 80 · #1

    • The real photograph was preferred in 2 of 3 blind comparisons

    • The real photograph was preferred in 3 of 3 blind comparisons

    • Even in 3 blind comparisons

    • Even in 3 blind comparisons

    • The real photograph was preferred in 3 of 3 blind comparisons

    • Even in 3 blind comparisons

When to use

Same brief, side by side

Prompt used:A Black woman with long locs, bare shoulders and a strapless cream top stands among dense tropical foliage and looks calmly into the lens. She holds an amber glass dropper bottle with a gold cap up in front of her chest with both hands, fingertips lightly touching it. Broad green leaves and tall reed stems fill the background; a blurred leaf crosses the bottom of the frame.

Preferred over the real photograph in 1 of 3 blind comparisons

View all 48 scenes →

Expert evaluation

Would use with caveats
The top rung we have: in blind tests its images were not reliably told apart from real photographs on believability. That is the ceiling of this scale, not proof of equality, but it is the model I would finish a hero image in.

Strengths

  • Not separated from real photographs on believability
  • Clearly ahead of every other model it was paired with
  • Widest lead over the open-weights floor, FLUX.2 Klein

Weaknesses

  • No measured lead over photographs: the score above 100 is the scale's ceiling
  • Close to even with photographs on believability

Rater agreement is low by the usual yardstick (Fleiss' kappa 0.16–0.20). Half the pairs compare adjacent tiers, where “equal” is an honest answer. Conclusions hold at the level of the whole corpus, not for a single scene, pair or rater.

Dmitry Shironosov

Dmitry Shironosov

Head of AI Research, Everypixel Workroom

What it costs

API list prices for GPT Image 2
ProviderPriceChecked
fal.aiMedium quality, 1024×1024$0.053 / imgSep 25, 2026
SegmindMedium quality, 1024×1024$0.06 / imgSep 25, 2026
WaveSpeedAIMedium quality, 1K$0.06 / imgSep 25, 2026

Prices are for one medium-quality 1024×1024 image, checked 25 September 2026; the price at the top of the page is the average of the providers listed. OpenAI itself prices GPT Image 2 by tokens rather than per image, so it is not in the table. Workroom pricing is on the pricing page.

Workroom pricing →

License and terms

GPT Image 2 is a proprietary OpenAI model available through the OpenAI API; its weights are not published. Terms for generated images are set by the service you generate them in.

Frequently asked questions

About Workroom

Workroom is an AI router: we give access to third-party models, including GPT Image 2, and we are not affiliated with OpenAI. We have no stake in any single vendor winning; the router is only useful if it sends each brief to the right model. The judgement-level data behind this card is published by Everypixel research under CC BY 4.0 and linked below, so every number here can be checked independently.

How we tested

Between 21 and 25 August 2026 we compared four rungs of our quality ladder blind on 48 commercial scenes: FLUX.2 Klein, Qwen Image 2, GPT Image 2 and the real photographs each scene was written from. Every rung was paired with every other on every scene — 288 pairs and 863 judgements — and each pair was judged three times, without any indication of which image came from which model, on three questions: which image follows the prompt better, which is more believable as a real photograph, and which is more attractive. In September GPT Image 2 was also paired blind with Seedream 4.0 on the same scenes: 432 more judgements.

How the scale is built

Pairwise comparisons are fitted with a Bradley–Terry model; intervals come from bootstrapping over scenes. The rungs of the scale are frozen from the baseline study, so every new model lands on the same scale. The score uses believability only:

score = 100 + (strength - strength_gold) / a

A real photograph is pinned at 100 by definition; slope 0.020835 strength per point; calibration points: floor, mid, gold. Category positions are fitted the same way on the 6 scenes of each category, with a real photograph pinned at 100 on every one of them.

Direct blind tests

The pairs GPT Image 2 was judged against directly — where its place on the scale comes from.

vs Qwen Image 2.0

GPT Image 2 was preferred over Qwen Image 2.0 in 72% of blind comparisons.

Gap (Bradley–Terry)
+0.83 (interval not published)
Preference
72.0% of blind comparisons preferred it (ties split, all questions, vote share)
Prompt adherence68 preferred · 53 equal · 23 other preferred
Believability93 preferred · 27 equal · 24 other preferred
Aesthetics104 preferred · 12 equal · 28 other preferred

144 blind comparisons, each pair shown 3 times

vs FLUX.2 Klein

GPT Image 2 was preferred over FLUX.2 Klein in 83% of blind comparisons.

Gap (Bradley–Terry)
+1.68 (interval not published)
Preference
83.3% of blind comparisons preferred it (ties split, all questions, vote share)
Prompt adherence102 preferred · 37 equal · 5 other preferred
Believability108 preferred · 27 equal · 9 other preferred
Aesthetics103 preferred · 30 equal · 11 other preferred

144 blind comparisons, each pair shown 3 times

vs Real photograph

Not separated from the real photograph by this data. Parity is not claimed.

Gap (Bradley–Terry)
+0.09 [−0.16, +0.33]
Preference
53.8% of blind comparisons preferred it (ties split, all questions, vote share)
Prompt adherence56 preferred · 56 equal · 32 other preferred
Believability42 preferred · 55 equal · 47 other preferred
Aesthetics57 preferred · 44 equal · 43 other preferred

144 blind comparisons, each pair shown 3 times

vs Qwen Image 2.0 Pro

GPT Image 2 was preferred over Qwen Image 2.0 Pro in 51% of blind comparisons.

Gap (Bradley–Terry)
+0.64 [+0.42, +0.86]
Preference
50.6% of blind comparisons preferred it (wins over all judgements, believability only, vote share)
Prompt adherence126 preferred · 246 equal · 59 other preferred
Believability218 preferred · 128 equal · 85 other preferred
Aesthetics235 preferred · 84 equal · 112 other preferred

431 blind comparisons, each pair shown 3 times

vs Seedream 4.0

GPT Image 2 was preferred over Seedream 4.0 in 69% of blind comparisons.

Gap (Bradley–Terry)
+0.78 [+0.55, +1.03]
Preference
68.6% of blind comparisons preferred it (ties split, all questions)
Prompt adherence187 preferred · 171 equal · 74 other preferred
Believability265 preferred · 94 equal · 73 other preferred
Aesthetics279 preferred · 52 equal · 101 other preferred

432 blind comparisons, each pair shown 3 times

Open data

Every image, every vote and every number behind the scores, so the conclusions can be checked independently.

Download: VotesCSV · 106 KB · CC BY 4.0Download: ResultsJSON · 5 KB · CC BY 4.0Download: Scenes and promptsCSV · 30 KB · CC BY 4.0Download: READMEMD · 2 KB · CC BY 4.0Download: ImagesZIP · 128 MB · All rights reserved — verification only

Data: CC BY 4.0. Images: All rights reserved; ten Unsplash photographs under the Unsplash License.

Read these before reusing the data

  • Prompt adherence is biased against the photograph: each prompt was written by looking at the photograph, and a model executes text literally, while a photograph always contains incidental detail the description does not mention. Raters count that as a miss — which is why the absolute scale is built on believability alone.
  • The share of ties is inflated: during collection the “equal” option was pre-selected, so a criterion the rater never touched was recorded as a genuine tie. The defect was fixed on 28 August 2026, after this data was collected. Tier ordering survives it: dropping every tie only widens the gaps.
  • All images were shown at 2048 px on the long edge — native output size for the generative tiers; the photographs were downscaled from 30–45 megapixels. On screen the photographs are not softer than the generated images.

Read the full study →

Cite this page

<blockquote cite="https://workroom.everypixel.com/models/gpt-image-2">
  <p>GPT Image 2 scores 104.4 on our photo-anchored scale, where a real photograph is fixed at 100.</p>
  <footer>&mdash; <a href="https://workroom.everypixel.com/models/gpt-image-2">GPT Image 2 benchmark: blind pairwise test</a>, Everypixel Workroom, August 2026</footer>
</blockquote>

Everypixel Workroom. (2026). GPT Image 2 benchmark: blind pairwise test. https://workroom.everypixel.com/models/gpt-image-2