Chinese characters come out wrong or invented
10 posters, EN / 中文 / mixed, text accuracy scored
We don't rank models in general. We test them one job at a time: the same inputs and settings for every model, graded blind, with the cost of each usable result. 5 of these jobs are being tested now.
Chinese characters come out wrong or invented
10 posters, EN / 中文 / mixed, text accuracy scored
Product details change; lighting looks fake
Same product photo → white background, lifestyle, on-model
Turnaround views don't match each other
1 description → front, side, back, expressions
Looks plastic or not like the person
5 consented source photos → headshots
Title text unreadable, layout cluttered
10 covers with headline text
Edits change things you didn't ask for
15 edit instructions on 5 photos
Generic marks, unusable raster files
5 briefs, vector output scored
Numbers and labels wrong
5 data briefs, accuracy scored
Style drift, anatomy errors
10 style briefs
Perspective and materials unrealistic
5 floor plans or sketches
The field as of 29 Oct 2026. Prices are list prices.
| Model | Latest version | List price | Mainland China |
|---|---|---|---|
| GPT Image 2.5 Sunburst / FlareOpenAI | 8 Sep 2026 | ~$0.21 at max; 0.5 Higgsfield credits at 1K medium | ? |
| Nano Banana 2.1Google | 6 Oct 2026 | ~$0.034 | ? |
| Seedream 5.0 ProByteDance | 12 Aug 2026 | from $0.045 | Yes |
| Midjourney V8.1 | 30 Apr 2026 | $10–120/mo | ? |
| FLUX 3 ImageBlack Forest Labs | announced 23 Jul 2026 | ? | ? |
| Ideogram 4.0 | Jun 2026 | ~$0.06 | ? |
| Recraft V4.1 | 14 May 2026 | $0.035 | ? |
| Qwen-Image 3.0Alibaba | ~Jul 2026 | ? | Likely |
How we choose and grade: our methodology.