October 28, 2026 | AI | By Mahesh
Haut.AI launched SkinFM on October 27, 2026, a skin foundation model that Haut.AI says was trained on more than 4 million real skin images and can assess more than 230 skin measurements and 48 biomarkers [1]. It arrived with SkinBench, a public benchmark built by Haut.AI, where the company reports 69.6 out of 100 against 50.6 for Gemini 3.5 Flash, the best of six general models tested [1][2]. Haut.AI ran the test itself, and no price or access route for SkinFM has been published.
- SkinFM is a single image model for skin analysis, and Haut.AI says it was trained on more than 4 million real skin images.
- Haut.AI scores 69.6 out of 100 on its own SkinBench against 50.6 for Gemini 3.5 Flash, and six general models flagged eye puffiness that was absent in 95% to 100% of images.
- SkinBench is public but SkinFM has no stated price or access route, and the benchmark is vendor-run and the release has inconsistent figures.
What SkinFM is and what is new
SkinFM is an image-based model that analyzes many skin characteristics with one network, in place of a separate model for each task [1]. Its measurements include wrinkles, redness and pigmentation [1]. Haut.AI says it builds brand-specific grading systems on top of the model for beauty companies [2].
The release centers on a claim about general-purpose AI. Haut.AI CEO and co-founder Anastasia Georgievskaya said that AI can recognize a face, but that does not mean it can accurately assess the skin [1]. The company argues that brands need evidence before they trust a vendor with consumer skin data.
Under the hood: what Haut.AI discloses
The disclosed details are thin. The release gives the 4 million image count, the 230 measurements and the 48 biomarkers, and it does not state the architecture, parameter count, training data sources or the skin tone mix of the training set [1]. That last gap matters because the benchmark includes an inclusivity score, which takes the result for the worst-performing skin tone group [2].
A repeatability figure of 0.97 to 0.98 intraclass correlation comes from a separate Haut.AI study of its measurement technology and was not produced by SkinBench [2]. SkinBench does not test repeatability yet, and Haut.AI says it plans to add that along with more models [1][2].
SkinBench: vendor-reported results
SkinBench has 19 tests across seven concerns: wrinkles, sun spots, redness, skin tone, skin type, dark circles and eye puffiness. Eleven tests grade features and eight measure how well a model locates them, with scores checked against expert reference assessments [1][2]. Haut.AI tested Gemini 3.5 Flash, GPT-5.6 Sol, Claude Fable 5, Grok 4.6, Qwen 3.8 27B and MedGemma 27B [1][2].
| Test | Haut.AI | Rival result |
|---|---|---|
| Overall Skin Intelligence Score | 69.6 | 50.6 (Gemini 3.5 Flash) |
| Skin type classification | 77.8 | 57.7 (strongest rival) |
| Frown line tracing, synthetic faces | 47.8 | 20.2 (GPT-5.6 Sol) |
| Deep forehead wrinkles found | 82% | 15% to 60% |
| Weakest skin tone group | 50.9 | 30.3 (next best general model) |
| Crow's feet | 61.8 | 71.7 (Gemini 3.5 Flash) |
The most striking result is false detection. When eye puffiness was absent, the six general models reported it in roughly 95% to 100% of images, against about 18% for Haut.AI [1][2]. Haut.AI also reports 0.82 agreement with experts on frown line severity against 0.73 for GPT-5.6 Sol and 0.86 for individual human graders, on 350 synthetic faces [1][2]. Gemini 3.5 Flash beat Haut.AI on crow's feet and on skin tone assessment, 90.4 to 87.6, so Haut.AI lost 2 of 19 tasks [1][2].
Pricing, availability and how to access it
SkinBench is public, with its leaderboard, comparisons and methodology at skinbench.ai [1]. SkinFM has no stated price, license, sign-up page or API, and the release describes it as technology for beauty brands [1][2]. Brands interested in it would need to contact Haut.AI directly.
Limitations, trade-offs and alternatives
Haut.AI built, ran and published the benchmark and also entered its own model, so the lead is a company claim and not an independent ranking. Cosmetics Business labels its article as promoted content [2]. The leaderboard counts each rival model's best result per test, and the general models were not built for skin grading [2]. The release also contradicts itself: the frown line scores of 0.82 for Haut.AI and 0.73 for GPT-5.6 Sol in the text differ from the figures in an image caption, so check skinbench.ai before quoting any single number [1].
For readers who build apps, the practical lesson is narrow. General chat models can over-report skin problems, and any consumer skin feature should be checked against expert-labeled images before launch. For another recent model release with disclosed specs, see Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks.
Sources
- Haut.AI. "Haut.AI Launches SkinFM, Its Skin Foundation Model, as New SkinBench Benchmark Exposes AI Skin Analysis Flaws," PR Newswire, October 27, 2026 (company press release). prnewswire.com
- Cosmetics Business. "SkinBench: six AI models saw skin problems that were not there" (labeled promoted content). cosmeticsbusiness.com
Read More on Depth Grid
- Nvidia in Talks to Deepen Stake in or Buy Reflection AI
- Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
- Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
- Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.