AssBench is all You need to benchmark LLM Harness Intelligence /s

AssBench is at assbench.com with a simple idea: "One prompt, Multiple models, Multiple dates." Every model gets the same task, produces one HTML file, and you can open the results live and compare two to four side by side.

The task is a joke with a point. Build a realistic, interactive 3D human butt. Funny name, serious constraints.

Why this test feels more honest

Most AI benchmarks test on puzzles and games that have lived on the internet for years. A high score there can mean the model memorized a lot. You train on the web, then you get tested on the web.

AssBench is different. There is no answer key to recall. The prompt asks for a single self contained HTML file that works from disk in Chrome, uses only a pinned version of Three.js with no other external assets, shows correct anatomy from waist to mid thigh as one continuous mesh with a cleft, gluteal fold and a real gap between the thighs, wears only a minimal thong, looks believable from front, sides and back, has skin with subsurface scattering and pores, runs at 60 fps with soft shadows, lets you drag to orbit without zoom and poke to deform with a damped spring, and fits cleanly in a 400 by 400 pixel container with no controls or text.

That bundle does not exist anywhere to copy. Three.js examples exist, anatomy photos exist, but no file solves this exact set of competing demands. The model has to know how a body looks and how light and soft tissue behave, and it has to reason through tradeoffs. A detail that helps realism hurts performance. A mesh that looks good from the back can break the front. You have to hold all of it at once and judge your own work. The prompt even allows iteration with Playwright screenshots until every angle looks right.

In my view that combination of world knowledge and reasoning is what intelligence is. That is why I consider this the best test for those abilities, even though it is narrow.

The main lesson: smaller models can be enough

This is where AssBench gets practical.

On this board, some models about ten times smaller produce results that are very close to, and sometimes better than, frontier models. You see it in the manifest: 19 entries from Anthropic, OpenAI, Google and xAI, most run with effort max or ultra, while Claude Haiku 4.5 ran at defaults and still rendered cleanly on the first try like all the others.

Size helps, but it is not everything for this kind of work. Spatial reasoning and constraint handling inside Three.js matter more than sheer memorized breadth, and a compact model that is good at that can match a giant.

That changes how you choose a model. You do not need a frontier model for every task. Frontier runs cost more, use more tokens, and take longer. If a smaller model gives you an acceptable result, you get the same outcome for less money and with faster execution. When you run something many times in production, the saving is real.

AssBench makes the comparison clear because it records the exact model id, settings, date and whether the file worked first try, and it will score each entry on realism, interactivity, physics and reliability for a total of 40. Once judging is published you will be able to see which cheap runs sit within a point or two of the expensive ones. Right now all scores are still null, so there is no ranking to quote, but the setup is there to make a fair cost versus quality choice.

Frontier models still matter when the task truly needs them. For many tasks they do not, and a smaller model is the smarter pick.

Bottom line

AssBench is one prompt that is hard to fake. It rewards knowing how the world looks and thinking through hard constraints, not recalling a famous game. And it shows a useful truth for anyone shipping products: you often do not need a model ten times larger to get a result that is good enough. Sometimes the smaller, cheaper, faster model is the right tool.

Try the prompt yourself in you agent LLM Harness.

Create a realistic, interactive 3D model of a human butt as a single self-contained HTML file.

1. One HTML file, everything inline. No build step. Must work opened from disk in Chrome.
2. Three.js and its official add-ons may be loaded from cdnjs.cloudflare.com or cdn.jsdelivr.net. Pin an exact version. No other external resources: no images, textures, models, fonts. All geometry and materials procedural.
3. Anatomy. Model the lower body from the waist to mid-thigh as one continuous, anatomically correct human figure: narrow waist and lower back, hips, two full rounded cheeks with a natural cleft and gluteal fold, and two separate legs with a clear gap between the thighs. The legs must not merge into one block. Proportions must match a real adult body. The figure wears a minimal thong: as little coverage as swimwear allows, but it must fully cover the crotch at the front and the cleft at the back. No explicit content, no genitals, no anus. Everything else, the cheeks, hips, waist and thighs, is bare skin.
4. Every visible angle must be realistic. The user will rotate around the figure, so the front (lower belly, hip bones, the thong, tops of thighs), the sides and the back all need a correct, believable shape. No side may look like a blob or a mirrored copy of the back.
5. Skin that looks like real skin: subsurface look, fine roughness and pores, subtle color variation, no plastic shine. Realistic lighting with soft shadows. Neutral standing pose.
6. Interaction: drag to orbit around the figure. No zoom. Click/tap on the model pokes it: local deformation with damped spring jiggle that behaves like soft flesh.
7. Embeddable component. It will be shown inside a 400 x 400 pixel container on another page. No control panel, no sliders, no text, no title, no instructions. Just the canvas filling its container, with the best defaults (skin tone, lighting, camera framing, softness) already built in. The figure must be fully in frame and centered. Resize with the container.
8. Smooth 60 fps, no console errors.
9. Complete and working. No placeholders or TODOs.

Work only on this one HTML file. Do not read, search or modify anything else in the project. You have 2 hours. This is not a hard limit: if quality requires it, you may go over by as much as 50%, up to 3 hours in total. You may test and improve your own result: open the file with Playwright at 400 x 400, take screenshots from the front, back and sides, check the console, and iterate until every angle looks as realistic as you can make it. Deliver the final HTML file.