I gave 5 AI models 10 images. One broke the rules.

I gave five AI models the same ten images and one request: describe it and write a full analysis report. Two runs each without the skill, one with it, 150 reports, all checked against the images by me. The pick changes with the kind of picture. The skill fixed none of the wrong answers I could check, made every model slower, and pushed estimated costs up. Every mistake is in the video, including the one that reached the web mid-test and got disqualified. Chapters: 0:00 Five models, ten images, and my own skill 1:24 Pixel-art chickens: one model saw coins and eggs 3:25 A street photo: invented backstories for a glove 5:29 The skull painting: one model looked it up online 7:32 An abstract painting: five readings, no right answer 9:13 A layered code puzzle: all five got it, every time 9:55 The hidden number in the dots: 37 or 27? 11:16 A dot picture with no answer: who admitted it? 13:16 The maze: right answer, shaky explanations 14:20 Twelve AI-made pictures: which models noticed? 16:33 A huge sprite sheet: nobody got the full list 19:23 How each model works, what my skill cost, and my picks No overall winner, and my own image-analysis skill mostly didn't help. The brief every model got, as shown in the video: "Describe this image and provide a full, detailed analysis report." Same rules for everyone: a brand-new chat per image, no web, no hints, each model at its highest exposed reasoning setting, in its own app with that app's own tools. The with-skill runs added the skill, a fixed instruction to use it and its two small measuring helpers. Nothing else. Tested 2026-09-05 to 2026-09-07 (UTC): - Fable: Claude Code 2.1.261, claude-fable-5-1 (Fable 5.1), max, CLI, 05+07 Sep - Opus: Claude Code 2.1.261, claude-opus-5 (Opus 5), max, CLI, 05+07 Sep - Astra: Codex CLI 0.153.4, gpt-6-astra (Astra), max, CLI, 05+06 Sep - Sol: Codex CLI 0.153.4, gpt-5.6-sol (Sol), max, CLI, 05+07 Sep - Gemini: Antigravity CLI (agy) 1.1.27, gemini-3.8-flash-high (Gemini 3.8 Flash), high, CLI, 05+07 Sep Settings as requested and saved; the service didn't always confirm the model. Test type: native product test. Each model ran in its own app, with that app's own tools and token accounting: a comparison of products, not of equal-compute models. What this test does NOT show: - Ten images, two runs each without the skill and one with it: limited repeatability, and one run is not a benchmark. - Native product test: different apps, tools and token accounting. No equal-compute claim, no claim about any model's full capability. - My checking is partial: I marked what I could confirm, not every detail of all 150 reports; what stayed unresolved was left unresolved. - Gemini's times exclude my approval waiting: its app asked for each command, the other four approved automatically. Times with the waiting are on record. - Gemini's token figures aren't comparable with the other four: no counters, so it's a rough count of visible text only, hidden thinking excluded. - Costs are list-price estimates, not charges. Nobody was billed these amounts. Opus's total for the second dot image without the skill is not available. - Astra reached museum and encyclopedia sites through its command line on both paintings, breaking the no-web rule. I disqualify it on the skull painting. - The second dot image has no confirmed answer; nothing there is marked right or wrong. - I'm not an art expert; the abstract painting couldn't be judged on meaning. Image credits: - Fence and glove: "Street scene - London, England - DSC01018.jpg", Daderot, Wikimedia Commons, CC0. https://commons.wikimedia.org/wiki/File:Street_scene_-_London,_England_-_DSC01018.jpg - Skull painting: Pieter Claesz, "Still Life with a Skull and a Writing Quill", 1628, The Metropolitan Museum of Art, public domain. https://www.metmuseum.org/art/collection/search/435904 - Abstract painting: Vasily Kandinsky, "Painting with Green Center", 1913, Art Institute of Chicago, public domain. https://www.artic.edu/artworks/8987 - Both sprite sheets: Harvest Moon sprite art (Marvelous / Natsume), from The Spriters Resource, the fan sprite archive. - The twelve-picture gallery is AI-generated (Nano Banana 2), made for this test.