The best models are close, and still imperfect.
Forty image models were asked to write Turkish words built around the dotless ı. Two AI reviewers scanned every image; a human reviewed a smaller sample, and the human verdict is the ground truth. It puts Nano Banana 2 first: about 92% of its reviewed images kept the requested text intact. GPT Image 2 and GPT Image 1.5 are close behind, so the top group matters more than the exact order.
Gemini 3.5 Flash and Claude Sonnet 4.6 both pick GPT Image 2 as the winner. GPT Image 1 Mini is surprisingly careful with the dot itself, wrongly dotting the ı only about 6% of the time in the human sample, but it trails the leaders on exact text and takes about 30 seconds per image.
Model rankings
Humans reviewed 1,055 of the 8,396 successful images; AI judges cover the rest.
Exact Dotted Substituted Mangled Off task
| 1 | 
Nano Banana 2
Google DeepMind (2026-02)
|
92%53 labels | 93% | 75% | 5%53 labels | |
$0.069 | 14.1s |
| 2 | 
GPT Image 2
OpenAI (2026-04)
|
91%23 labels | 97% | 81% | 5%23 labels | |
$0.0529 | 54.3s |
| 3 |
GPT Image 1.5
OpenAI (2025-12)
|
91%53 labels | 84% | 78% | 4%53 labels | |
$0.087 | 25.3s |
| 4 |
Nano Banana Pro
Google (2025-11)
|
75%20 labels | 87% | 70% | 18%20 labels | |
$0.138 | 21.0s |
| 5 |
GPT Image 1
OpenAI (2025-03)
|
75%20 labels | 82% | 75% | 19%20 labels | |
$0.167 | 26.0s |
| 6 | 
Grok Imagine Image Quality
xAI (2026-04)
|
74%23 labels | 78% | 73% | 20%23 labels | |
$0.05 | 5.4s |
| 7 | 
FLUX.2 [pro]
Black Forest Labs (2025-11)
|
72%53 labels | 62% | 63% | 19%53 labels | |
$0.03 | 8.9s |
| 8 |
GPT Image 1 Mini
OpenAI (2025-04)
|
70%20 labels | 83% | 75% | 6%20 labels | |
$0.036 | 30.2s |
| 9 | 
Krea 2 Large
Krea AI (2026-05)
|
70%23 labels | 59% | 58% | 22%23 labels | |
$0.06 | 25.4s |
| 10 | 
Seedream 5.0 Lite
ByteDance Seed (2026-02)
|
65%23 labels | 48% | 52% | 26%23 labels | |
$0.035 | 30.9s |
| 11 |
Krea 2 Medium
Krea (2026-05)
|
60%20 labels | 63% | 65% | 35%20 labels | |
$0.03 | 12.9s |
| 12 |
FLUX.2 [flex]
Black Forest Labs (2025-11)
|
55%20 labels | 66% | 61% | 29%20 labels | |
$0.06 | 8.5s |
| 13 |
FLUX.2 [dev]
Black Forest Labs (2025-11)
|
51%53 labels | 51% | 56% | 41%53 labels | |
$0.0101 | 4.9s |
| 14 | 
Ideogram 4.0
Ideogram (2026-06)
|
50%20 labels | 75% | 76% | 21%20 labels | |
$0.06 | 34.4s |
| 15 | 
Z-Image Turbo
Alibaba Tongyi-MAI (2025-11)
|
48%23 labels | 25% | 26% | 25%23 labels | |
$0.0006 | 3.3s |
| 16 |
Imagen 4
Google DeepMind (2025-05)
|
43%23 labels | 55% | 54% | 31%23 labels | |
$0.04 | 7.2s |
| 17 |
Imagen 4 Ultra
Google (2025-06)
|
40%20 labels | 50% | 53% | 31%20 labels | |
$0.06 | 20.6s |
| 18 |
Nano Banana
Google (2025-08)
|
40%20 labels | 32% | 39% | 40%20 labels | |
$0.039 | 5.9s |
| 19 |
Recraft V4.1
Recraft (2026-05)
|
40%20 labels | 34% | 47% | 56%20 labels | |
$0.04 | 6.8s |
| 20 | 
Qwen-Image-2.0
Alibaba (2026-02)
|
35%20 labels | 28% | 30% | 44%20 labels | |
$0.035 | 6.5s |
| 21 |
Qwen-Image-2.0-Pro
Alibaba (2026-02)
|
35%20 labels | 31% | 30% | 50%20 labels | |
$0.075 | 11.8s |
| 22 |
Imagen 4 Fast
Google (2025-06)
|
30%20 labels | 35% | 35% | 25%20 labels | |
$0.02 | 5.0s |
| 23 |
FLUX.2 [klein] 9B
Black Forest Labs (2026-01)
|
30%20 labels | 10% | 11% | 43%20 labels | |
$0.003 | 5.4s |
| 24 |
Ideogram 2a
Ideogram AI (2025-02)
|
26%23 labels | 10% | 11% | 53%23 labels | |
$0.04 | 12.1s |
| 25 |
Seedream 4.0
ByteDance Seed (2025-09)
|
22%23 labels | 12% | 18% | 75%23 labels | |
$0.03 | 8.1s |
| 26 |
Seedream 4.5
ByteDance Seed (2025-12)
|
19%53 labels | 18% | 25% | 77%53 labels | |
$0.04 | 13.1s |
| 27 |
FLUX1.1 [pro]
Black Forest Labs (2024-10)
|
17%23 labels | 6% | 8% | 25%23 labels | |
$0.04 | 4.3s |
| 28 |
Ideogram 3.0
Ideogram AI (2025-05)
|
15%53 labels | 9% | 18% | 62%53 labels | |
$0.06 | 9.0s |
| 29 |
FLUX.1 [dev]
Black Forest Labs (2024-08)
|
13%23 labels | 4% | 4% | 43%23 labels | |
$0.0039 | 4.3s |
| 30 |
Qwen-Image
Alibaba Qwen (2025-08)
|
13%23 labels | 7% | 7% | 53%23 labels | |
$0.0052 | 13.1s |
| 31 |
Imagen 3
Google DeepMind (2024-05)
|
13%23 labels | 7% | 4% | 54%23 labels | |
$0.04 | 6.2s |
| 32 |
Qwen-Image-2512
Alibaba Qwen (2025-12)
|
13%23 labels | 10% | 13% | 71%23 labels | |
$0.005 | 8.8s |
| 33 |
Ideogram 2.0
Ideogram AI (2024-08)
|
13%23 labels | 7% | 16% | 80%23 labels | |
$0.08 | 21.0s |
| 34 | 
Stable Diffusion 3
Stability AI (2025-06)
|
10%20 labels | 2% | 3% | 29%20 labels | |
$0.0013 | 3.9s |
| 35 |
Imagen 3 Fast
Google (2024-07)
|
10%20 labels | 6% | 6% | 63%20 labels | |
$0.02 | 7.0s |
| 36 |
FLUX.1 [schnell]
Black Forest Labs (2024-08)
|
9%23 labels | 4% | 7% | 38%23 labels | |
$0.0013 | 3.0s |
| 37 | 
HiDream-I1 Dev
HiDream.ai (2025-04)
|
9%23 labels | 6% | 6% | 59%23 labels | |
$0.0045 | 9.9s |
| 38 |
Ideogram 1.0
Ideogram AI (2024-02)
|
4%23 labels | 2% | 8% | 43%23 labels | |
$0.06 | 17.3s |
| 39 | 
Stable Diffusion 1.5
Runway ML and Stability AI (2022-10)
|
0%23 labels | 0% | 0% | – | |
$0.001 | 3.9s |
| 40 |
SDXL 1.0 base
Stability AI (2023-07)
|
0%23 labels | 2% | 1% | 33%23 labels | |
$0.0019 | 5.4s |
Cost curves
The winners do not ask for the same tradeoff. Nano Banana 2 leads the human review at about $0.069 per image and 14 seconds. GPT Image 2 costs less, about $0.053, but it makes you wait about 54 seconds.
Cheap models start to work, then the ceiling gets expensive
FLUX.2 makes the first jump into usable territory. The highest human scores sit farther right: newer GPT Image and Gemini-family models buy better letters with more money, more waiting, or both.
The full-run AI review makes GPT Image 2 stand out
Gemini 3.5 Flash reviews every successful image and puts GPT Image 2 on top. GPT Image 1 Mini is the interesting cheaper pocket: not the winner, but surprisingly strong for the price.
Release and operating tradeoffs
Model releases are moving in the right direction. The best recent systems are clustered close together, so the practical question becomes which tradeoff you can live with: cost, waiting time, or a few more dotted letters.
Image models got much better over time
The trend is clear: newer models usually keep more Turkish text intact. GPT Image 1.5, Nano Banana 2, and GPT Image 2 sit close together at the top.
Waiting time is its own price
GPT Image 2 is strong and cheaper than Nano Banana 2, but it is much slower. Grok is fast; GPT Image 1 Mini is a cheaper middle ground with a longer wait.
What the AI judges saw
AI reviewers looked at many more images than the human evaluation covered. They agree on the broad shape: GPT Image and Nano Banana models stay near the top, and GPT Image 2 leads both AI reviews.
Gemini 3.5 Flash
| Rank | Model | Reviewed | Exact text | Dotted ı rate | Failure mix |
|---|---|---|---|---|---|
| 1 |
GPT Image 2
OpenAI (2026-04)
|
210 | 97% | 2% | |
| 2 |
Nano Banana 2
Google DeepMind (2026-02)
|
210 | 93% | 5% | |
| 3 |
Nano Banana Pro
Google (2025-11)
|
210 | 87% | 6% | |
| 4 |
GPT Image 1.5
OpenAI (2025-12)
|
210 | 84% | 11% | |
| 5 |
GPT Image 1 Mini
OpenAI (2025-04)
|
207 | 83% | 8% | |
| 6 |
GPT Image 1
OpenAI (2025-03)
|
210 | 82% | 10% | |
| 7 |
Grok Imagine Image Quality
xAI (2026-04)
|
210 | 78% | 18% | |
| 8 |
Ideogram 4.0
Ideogram (2026-06)
|
210 | 75% | 14% | |
Claude Sonnet 4.6
| Rank | Model | Reviewed | Exact text | Dotted ı rate | Failure mix |
|---|---|---|---|---|---|
| 1 |
GPT Image 2
OpenAI (2026-04)
|
210 | 81% | 10% | |
| 2 |
GPT Image 1.5
OpenAI (2025-12)
|
210 | 78% | 14% | |
| 3 |
Ideogram 4.0
Ideogram (2026-06)
|
210 | 76% | 15% | |
| 4 |
GPT Image 1 Mini
OpenAI (2025-04)
|
207 | 75% | 14% | |
| 5 |
Nano Banana 2
Google DeepMind (2026-02)
|
210 | 75% | 14% | |
| 6 |
GPT Image 1
OpenAI (2025-03)
|
210 | 75% | 14% | |
| 7 |
Grok Imagine Image Quality
xAI (2026-04)
|
210 | 73% | 22% | |
| 8 |
Nano Banana Pro
Google (2025-11)
|
210 | 70% | 11% | |
Gemini 3 Pro
| Rank | Model | Reviewed | Exact text | Dotted ı rate | Failure mix |
|---|---|---|---|---|---|
| 1 |
GPT Image 2
OpenAI (2026-04)
|
23 | 87% | 5% | |
| 2 |
Nano Banana 2
Google DeepMind (2026-02)
|
53 | 87% | 5% | |
| 3 |
GPT Image 1.5
OpenAI (2025-12)
|
53 | 75% | 7% | |
| 4 |
Grok Imagine Image Quality
xAI (2026-04)
|
23 | 70% | 20% | |
| 5 |
GPT Image 1 Mini
OpenAI (2025-04)
|
20 | 65% | 18% | |
| 6 |
GPT Image 1
OpenAI (2025-03)
|
20 | 65% | 19% | |
| 7 |
FLUX.2 [pro]
Black Forest Labs (2025-11)
|
53 | 60% | 28% | |
| 8 |
Nano Banana Pro
Google (2025-11)
|
20 | 60% | 18% | |
GPT-5.5
| Rank | Model | Reviewed | Exact text | Dotted ı rate | Failure mix |
|---|---|---|---|---|---|
| 1 |
Nano Banana 2
Google DeepMind (2026-02)
|
52 | 90% | 5% | |
| 2 |
GPT Image 2
OpenAI (2026-04)
|
23 | 87% | 5% | |
| 3 |
Grok Imagine Image Quality
xAI (2026-04)
|
22 | 82% | 16% | |
| 4 |
GPT Image 1.5
OpenAI (2025-12)
|
53 | 81% | 7% | |
| 5 |
Nano Banana Pro
Google (2025-11)
|
19 | 74% | 13% | |
| 6 |
GPT Image 1
OpenAI (2025-03)
|
19 | 68% | 19% | |
| 7 |
GPT Image 1 Mini
OpenAI (2025-04)
|
19 | 68% | 25% | |
| 8 |
Krea 2 Large
Krea AI (2026-05)
|
23 | 61% | 22% | |
Claude Opus
| Rank | Model | Reviewed | Exact text | Dotted ı rate | Failure mix |
|---|---|---|---|---|---|
| 1 |
GPT Image 2
OpenAI (2026-04)
|
23 | 87% | 5% | |
| 2 |
Nano Banana 2
Google DeepMind (2026-02)
|
53 | 79% | 11% | |
| 3 |
Grok Imagine Image Quality
xAI (2026-04)
|
23 | 65% | 32% | |
| 4 |
GPT Image 1.5
OpenAI (2025-12)
|
53 | 64% | 17% | |
| 5 |
GPT Image 1 Mini
OpenAI (2025-04)
|
19 | 63% | 19% | |
| 6 |
FLUX.2 [pro]
Black Forest Labs (2025-11)
|
53 | 62% | 26% | |
| 7 |
Krea 2 Large
Krea AI (2026-05)
|
23 | 48% | 22% | |
| 8 |
Seedream 5.0 Lite
ByteDance Seed (2026-02)
|
23 | 48% | 32% | |
Qwen-VL
| Rank | Model | Reviewed | Exact text | Dotted ı rate | Failure mix |
|---|---|---|---|---|---|
| 1 |
Nano Banana 2
Google DeepMind (2026-02)
|
44 | 98% | 0% | |
| 2 |
GPT Image 1.5
OpenAI (2025-12)
|
45 | 96% | 0% | |
| 3 |
Seedream 4.5
ByteDance Seed (2025-12)
|
45 | 87% | 0% | |
| 4 |
Imagen 4
Google DeepMind (2025-05)
|
15 | 87% | 9% | |
| 5 |
Ideogram 3.0
Ideogram AI (2025-05)
|
45 | 84% | 3% | |
| 6 |
FLUX.2 [pro]
Black Forest Labs (2025-11)
|
45 | 84% | 5% | |
| 7 |
FLUX.1 [dev]
Black Forest Labs (2024-08)
|
15 | 80% | 0% | |
| 8 |
Seedream 4.0
ByteDance Seed (2025-09)
|
15 | 80% | 0% | |
See Hugging Face Dataset for the underlying rows, schema, and reusable benchmark splits. The leaderboard here is the readable site view; the Dataset is the canonical package for analysis and citation.
Prompt effects
Prompt wording changes the kind of failure more than it solves the problem. English coaching nudges exact text up to 46%, from 41% for the plain studio sign, but handwritten prompts dot the ı most often. Turkish prompts cut wrong dots sharply, then lose many images to unreadable or off-task output.
| Context | Language | Glyph guidance | Reviewed | Exact text | Dotted among legible text | Readable text |
|---|---|---|---|---|---|---|
| Coached sign prompt | English | yes | 130 | 46% | 38% | 96% |
| Turkish sign prompt | Turkish | no | 132 | 42% | 23% | 73% |
| Serif poster | English | no | 72 | 42% | 30% | 92% |
| Studio sign | English | no | 443 | 41% | 38% | 90% |
| Turkish coached sign prompt | Turkish | yes | 129 | 41% | 28% | 77% |
| LED matrix | English | no | 79 | 38% | 46% | 86% |
| Handwritten note | English | no | 70 | 36% | 49% | 97% |