---
title: "GPT Image 2 Blind Test Results: Scores by Use Case"
canonical_url: "https://workroom.everypixel.com/models/gpt-image-2"
score: 104.4
rank: 1
scale: "Anchor comparison of generative image models, v3"
tested: 2026-08-25
revised: 2026-09-14
---

# GPT Image 2 Blind Test Results: Scores by Use Case

**Score:** 104.4, above the scale's ceiling. A real photograph is 100.
**Rank:** #1 on the scale
**Author:** Dmitry Shironosov (Head of AI Research, Everypixel Workroom)
**Tested:** 2026-08-25 · revised 2026-09-14
**Canonical URL:** https://workroom.everypixel.com/models/gpt-image-2

## Independence

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.

## TL;DR

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.

## Where it stands

| Rank | Model | Score | 95% interval |
|---|---|---|---|
| #1 | GPT Image 2 (this card) | 104 | — |
| — | Real photograph | 100 | — |
| #2 | Qwen Image 2.0 Pro | 80 | 68.6–93.8 |
| #3 | Qwen Image 2.0 | 65 | — |
| #4 | Seedream 4.0 | 57 | 44–70 |
| #5 | FLUX.2 Klein | 24 | — |

Axis: believability. A real photograph is 100. A score above 100 is the scale's technical ceiling and does not mean parity.

### By use case

| Category | Score | Rank |
|---|---|---|
| Fashion and beauty | 117 | #1 |
| Interior | 114 | #2 |
| Street and architecture | 114 | #1 |
| Social marketing | 112 | #1 |
| Product and e-commerce | 112 | #1 |
| Food and beverage | 91 | #1 |
| Layout and typography | 85 | #1 |
| People and lifestyle | 80 | #1 |

*Category positions come from 6 scenes each — a tendency, not a measurement. A real photograph is 100 on every ruler.*

## When to use as default

Use GPT Image 2 when the image has to pass as a photograph: it was the only model in our test whose believability was not separated from the reference photographs, and it beat every other model it was paired with.

## When to use as finisher

Use it as the finisher after drafting in a mid-tier model: against Qwen Image 2.0 and Seedream 4.0 it was preferred in 72.0% and 68.6% of judgements, so the final hero asset is where its lead pays off.

## When to look elsewhere

Do not read 104 as better than a photograph: above 100 the scale has no room to separate models from the reference. Against the photographs its edge came from prompt adherence and aesthetics — and prompt adherence is biased against photographs — while believability was close to even.

## Direct blind tests

| Opponent | Result | Gap (Bradley–Terry, 95% CI) | Preferred in | Judgements |
|---|---|---|---|---|
| Qwen Image 2.0 | GPT Image 2 was preferred over Qwen Image 2.0 in 72% of blind comparisons. | +0.83 (interval not published) | 72.0% (ties split, all questions, vote share) | 144 |
| FLUX.2 Klein | GPT Image 2 was preferred over FLUX.2 Klein in 83% of blind comparisons. | +1.68 (interval not published) | 83.3% (ties split, all questions, vote share) | 144 |
| Real photograph | Not separated from the real photograph by this data. Parity is not claimed. | +0.09 [−0.16, +0.33] | 53.8% (ties split, all questions, vote share) | 144 |
| Qwen Image 2.0 Pro | GPT Image 2 was preferred over Qwen Image 2.0 Pro in 51% of blind comparisons. | +0.64 [+0.42, +0.86] | 50.6% (wins over all judgements, believability only, vote share) | 431 |
| Seedream 4.0 | GPT Image 2 was preferred over Seedream 4.0 in 69% of blind comparisons. | +0.78 [+0.55, +1.03] | 68.6% (ties split, all questions) | 432 |

## Price

| Provider | Price | Checked |
|---|---|---|
| [fal.ai](https://fal.ai/models/openai/gpt-image-2) | $0.053 / img | 2026-09-25 |
| [Segmind](https://www.segmind.com/models/gpt-image-2/pricing) | $0.06 / img | 2026-09-25 |
| [WaveSpeedAI](https://wavespeed.ai/models/openai/gpt-image-2/text-to-image) | $0.06 / img | 2026-09-25 |

Average: $0.058 / img across 3 providers

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.

## License

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.

## FAQ

### Is GPT Image 2 better than a real photograph?

This data makes no such claim. Its believability gap to the reference photographs is +0.09 with a 95% interval from −0.16 to +0.33, which includes zero, so the two are not separated. The 104.4 score reflects the scale's ceiling: a photograph is fixed at 100 and the scale has no room above it to measure a lead.

### Why was GPT Image 2 left out of the scale's calibration?

A calibration point has to be separated from its neighbour, otherwise the slope of the scale becomes arbitrary. Because GPT Image 2 was not separated from the photographs, the scale was fitted on FLUX.2 Klein, Qwen Image 2 and the photographs, and GPT Image 2's score was then read off that line.

### Is GPT Image 2 better than Seedream 4.0?

Clearly ahead. In the September session the believability gap was +0.78 with a 95% interval from +0.55 to +1.03, which does not include zero, and GPT Image 2 was preferred in 68.6% of 864 judgements with ties split. On the photo-anchored scale that is 104 against 57.

### How was GPT Image 2 tested?

It was one of four rungs compared blind on 48 commercial scenes in August 2026, paired on every scene with FLUX.2 Klein, Qwen Image 2 and the reference photographs, each pair judged three times on three questions. In September it was also paired with Seedream 4.0 on the same scenes.

### Is Workroom affiliated with OpenAI?

No. Workroom is an AI router that gives access to third-party models, including GPT Image 2, and is not affiliated with OpenAI. The judgement-level data behind this card is published by Everypixel research under CC BY 4.0, so every number on this page can be checked against the source dataset.

## Methodology

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.

- Scenes: 48; generations per scene: 1
- Compared with: FLUX.2 Klein, Qwen Image 2.0, Real photograph
- Pairs: 288; judgements: 863; each pair judged 3 times
- Collected: 2026-08-21 … 2026-08-25
- Questions: “Which image follows the prompt better?” “Which is more believable as a real photograph?” “Which is more attractive?”
- Scale: 0.020835 strength per point; A real photograph is 100.
- Category positions are fitted from the same judgements, 6 scenes per category, on the same scale.

### Caveats

- 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.

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.

## Links

- Full study: https://research.everypixel.com/anchor-ladder/
- Votes (CSV, 106 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/votes.csv
- Results (JSON, 5 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/results.json
- Scenes and prompts (CSV, 30 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/scenes.csv
- README (MD, 2 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/README.md
- Images (ZIP, 128 MB, All rights reserved — verification only): https://wr.everypixel.com/ghost-research/geo-benchmarks/datasets/anchor-comparison-v3-images.zip

## Citation

Everypixel research (2026). Anchor comparison of generative image models, v3 [Data set]. Ratings collected 21–25 August 2026.

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

---
HTML version: https://workroom.everypixel.com/models/gpt-image-2
Last modified: 2026-09-14