We compared three models on three scenes of Japanese businesswomen: a meeting-room presentation, a cafe laptop scene and an office PC scene. GPT Image 2 achieved the highest average score (3.93) and was rated usable in every scene. FLUX.2 [dev] was the cheapest at about $0.012 per image, but it drew an Apple logo on a laptop in the cafe scene, making that output commercially unusable. NanoBanana 2 was usable across all scenes but scored slightly lower.
Results
| Model | Avg Score | Wins | Cost / image | Verdict |
|---|---|---|---|---|
| GPT Image 2 | 3.93 | 3 | $0.0800 | Usable in all scenes |
| FLUX.2 [dev] | 3.75 | 0 | $0.0120 | Cheapest; trademark risk in cafe scene |
| NanoBanana 2 | 3.50 | 0 | $0.0800 | Usable but lower score |
Key takeaways
- GPT Image 2 was the most consistent across office, meeting and cafe scenes with an average score of 3.93.
- FLUX.2 [dev] is about 85% cheaper per image, but generated an Apple logo on a laptop, creating trademark risk for commercial use.
- NanoBanana 2 produced usable images in every scene but did not win any prompt outright.
- Cost per usable image can be misleading when a low-cost output fails commercial usability.
Verdict
For Japanese businesswomen, GPT Image 2 is the safest choice for quality and commercial usability; FLUX.2 [dev] is best for low-cost iteration with careful trademark review; NanoBanana 2 is a stable fallback.
Note
This is an English summary of the full Japanese benchmark article. Methodology, prompts, all outputs and cost details are documented in the original post.
Read the full article in Japanese →