7 skin parameters · 19 tasks · 7 AI models
The AI benchmark for skincare, beauty and dermatology
SkinBench tests how accurately AI models measure skin. Haut.AI and other AI models are scored on the same images against ground truth, parameter by parameter.
Illustration of SkinBench findings
Leaderboard
Every model scored on the same skincare tasks. The Skin Intelligence Score is the average of seven skin parameter scores. 100 means perfect agreement with expert ground truth on every image.
Skin Intelligence Score
AI model scores out of 100, higher is better.
Haut.AI
Gemini 3.5 Flash
GPT-5.6 Sol
Claude Fable 5
Qwen3.8 27B
Grok 4.6
MedGemma 27BHaut.AI SkinBenchClass prediction
AI model scores out of 100, higher is better.
Haut.AIHaut.AIGemini 3.5Gemini 3.5 FlashClaude 5Claude Fable 5GPT-5.6GPT-5.6 SolQwen3.8Qwen3.8 27BGrok 4.6Grok 4.6MedGemmaMedGemma 27BHaut.AI SkinBenchSegmentation
AI model scores out of 100, higher is better.
Haut.AIHaut.AIGemini 3.5Gemini 3.5 FlashGPT-5.6GPT-5.6 SolGrok 4.6Grok 4.6Claude 5Claude Fable 5MedGemmaMedGemma 27BQwen3.8Qwen3.8 27BHaut.AI SkinBenchInclusivity
AI model scores out of 100, higher is better.
Haut.AIHaut.AIGemini 3.5Gemini 3.5 FlashClaude 5Claude Fable 5GPT-5.6GPT-5.6 SolQwen3.8Qwen3.8 27BGrok 4.6Grok 4.6MedGemmaMedGemma 27BHaut.AI SkinBench| Rank | Model | Skin Intelligence Score | Class prediction | Segmentation | Inclusivity |
|---|---|---|---|---|---|
| 1 | Haut.AI | 69.6 | 75.9 | 46.9 | 50.9 |
| 2 | Gemini 3.5 FlashGoogle | 50.6 | 60.0 | 17.4 | 30.3 |
| 3 | GPT-5.6 SolOpenAI | 45.9 | 56.4 | 15.2 | 23.0 |
| 4 | Claude Fable 5Anthropic | 45.4 | 59.0 | 10.7 | 24.0 |
| 5 | Qwen3.8 27BAlibaba | 41.7 | 55.0 | 2.6 | 17.2 |
| 6 | Grok 4.6xAI | 39.2 | 47.6 | 11.4 | 16.8 |
| 7 | MedGemma 27BGoogle | 28.5 | 31.2 | 3.0 | 6.5 |
Haut.AI SkinBenchEach chart is ranked by its own metric. Higher is better for all scores. Each benchmark scores 100 × metric × share of images the model answered; a benchmark with no result scores 0. Consistency of predictions is coming in version 2.
How the scores are built
Metrics: QWK for grades, balanced accuracy for skin type, clDice for line tracing, IoU for dark circle and puffiness tracing, boundary-tolerant Dice (10 px) for sun spot tracing. Benchmarks of the same task on different datasets are averaged into one task score. Class prediction is the mean of the 11 grading tasks and Segmentation the mean of the 8 tracing tasks. Each skin parameter is the mean of its tasks (Wrinkles covers all line grading and tracing tasks), and the Skin Intelligence Score is the mean of the seven parameters.
Inclusivity is the score on the model's worst skin tone group: light (MST 1 to 4, Haut.AI scale 1 to 3), medium (MST 5 to 6, Haut.AI 4 to 5) or dark (MST 7 to 10, Haut.AI 6 to 7). On each benchmark with at least two groups of 30 or more images, each group scores 100 × metric × share answered within that group, and the lowest group counts. Results are averaged across datasets, then across tasks. Skin tone benchmarks and datasets without skin tone data are excluded.
Test your brand’s dataset against Haut.AI’s SkinFM
Scores by skin parameter
How each model scores on the seven skin parameters, out of 100.
Wrinkles
AI model scores out of 100, higher is better
Crow's feet, forehead, glabellar, lip and under-eye lines, graded and traced.
Haut.AIGemini 3.5 FlashGPT-5.6 SolClaude Fable 5Qwen3.8 27BGrok 4.6MedGemma 27B
Haut.AI SkinBenchSun spots
AI model scores out of 100, higher is better
Haut.AIClaude Fable 5GPT-5.6 SolGemini 3.5 FlashGrok 4.6Qwen3.8 27BMedGemma 27B
Haut.AI SkinBenchRedness
AI model scores out of 100, higher is better
Haut.AIHaut.AIGemini 3.5Gemini 3.5 FlashQwen3.8Qwen3.8 27BGPT-5.6GPT-5.6 SolClaude 5Claude Fable 5Grok 4.6Grok 4.6MedGemmaMedGemma 27BHaut.AI SkinBenchSkin tone
AI model scores out of 100, higher is better
Gemini 3.5Gemini 3.5 FlashClaude 5Claude Fable 5Grok 4.6Grok 4.6Haut.AIHaut.AIQwen3.8Qwen3.8 27BGPT-5.6GPT-5.6 SolMedGemmaMedGemma 27BThe one parameter where other models lead, by less than 3 points. It is also the most general task: identifying a tone, not measuring a skin condition.
Haut.AI SkinBenchSkin type
AI model scores out of 100, higher is better
Haut.AIHaut.AIGPT-5.6GPT-5.6 SolQwen3.8Qwen3.8 27BGemini 3.5Gemini 3.5 FlashClaude 5Claude Fable 5Grok 4.6Grok 4.6MedGemmaMedGemma 27BHaut.AI SkinBenchDark circles
AI model scores out of 100, higher is better
Haut.AIGemini 3.5 FlashGPT-5.6 SolClaude Fable 5Grok 4.6Qwen3.8 27BMedGemma 27B
Haut.AI SkinBenchEye puffiness
AI model scores out of 100, higher is better
Haut.AIGemini 3.5 FlashClaude Fable 5Grok 4.6Qwen3.8 27BGPT-5.6 SolMedGemma 27B
Haut.AI SkinBenchThe model behind the scores
Meet SkinFM
Haut.AI’s skin foundation model
Our skin analysis is powered by SkinFM, trained on 20 million images. One foundation replaces a separate model for every task, so we can build new measurements faster and tailor grading to each brand.
SkinBench puts it to the test against six other AI models.
Built to test how well AI measures skin
Brands and consumers can now ask AI models about skin. SkinBench measures how well they do it, on the same images and against the same expert ground truth as Haut.AI.
Wrinkles, sun spots, redness, skin tone, skin type, dark circles and eye puffiness.
11 grading tasks and 8 tracing tasks, every model scored on the same images.
Seven parameter scores averaged into one number out of 100.
Every skin parameter score behind the overall score.
What the results show
A model can return a plausible score without looking at the right place. Segmentation shows where each model actually looked.
On tracing wrinkles, the best other model scores less than half of what Haut.AI scores.
Grading is where other models come closest. Haut.AI leads on four of the five line grading tasks.
A model can score well overall and still fail on less common cases, such as the darkest skin tones. An average alone won't show it.
Want this accuracy in your product?
See how Haut.AI measures skin on your own images.
SkinBench: frequently asked questions
What is SkinBench?
SkinBench is Haut.AI’s benchmark for AI skin measurement. It compares Haut.AI with six other AI models on seven skin parameters, using identical test sets, ground truth and published metric definitions, and ranks them with one Skin Intelligence Score.
Which models are in the benchmark?
The benchmark compares seven models: Haut.AI, Gemini 3.5 Flash (Google), GPT-5.6 Sol (OpenAI), Claude Fable 5 (Anthropic), Qwen3.8 27B (Alibaba), Grok 4.6 (xAI) and MedGemma 27B (Google). We test more models and tasks internally. The public board shows only the results where every listed model was tested on every listed task, so all scores are directly comparable.
How is SkinBench different from clinical validation?
Clinical validation tests one model in a clinical study, against dermatologists and established dermatological scales. SkinBench scores many models on the same test sets against the same ground truth, so they can be compared directly. They answer different questions.
Does Haut.AI lose on any task?
Yes, on 2 of the 19 tasks. Skin tone: Gemini 3.5 Flash (90.4), Claude Fable 5 (88.7) and Grok 4.6 (87.9) all score above Haut.AI’s 87.6, each within 3 points. Crow’s feet grading: Gemini 3.5 Flash (71.7) and Claude Fable 5 (64.2) score above Haut.AI’s 61.8; on crow’s feet tracing, Haut.AI leads 60.4 to 25.3 (Gemini 3.5 Flash).
Is the gap statistically significant?
Yes. On every board we tested statistically (16 boards, paired bootstrap over images), the difference between Haut.AI and the other models is significant (p ≤ 0.01).
How often is SkinBench updated?
Usually quarterly, and whenever a major model update or a notable new AI model is released. An automated pipeline re-runs every model on the same test sets. Haut.AI updates its own models monthly.
Get in touch
Have questions or want to explore a partnership?
Fill out the form and our team will get back to you shortly.