BeckonBench
Each prompt is frozen for the season and pasted into a fresh session. One shot, no follow-ups. What you see on the leaderboard is the first thing the model said back.
Measures creative coding, game logic, atmosphere.
Build a playable haunted-house horror game as a single self-contained HTML file (no external assets, libraries, or network requests). The player explores a dark mansion, something is hunting them, and there is a way to win and a way to die. Prioritize atmosphere: darkness, lighting, sound (Web Audio API), and tension. Include brief on-screen instructions. It must work by simply opening the file in a browser.
Measures physics intuition, animation quality, patience with subtlety.
Create a realistic lava lamp simulation as a single self-contained HTML file (no external libraries or assets). Blobs of wax heat up at the bottom, rise, cool at the top, sink, and they deform, merge, and split like real wax — not rigid circles. Get the slow, hypnotic pacing right: it should be mesmerizing to watch for a full minute. It must work by simply opening the file in a browser.
Measures design taste, copywriting, responsive layout.
Design a landing page for a fictional product called "Nightjar" — an app that turns your rambling voice memos into organized, searchable notes. Single self-contained HTML file: no external assets, fonts, or libraries (system font stacks are fine). Write real marketing copy — no lorem ipsum, no placeholder text. It must look great at desktop width and on a phone. Aim for a visual style you would defend in front of a design director.
Measures spatial reasoning in code, ambition, personality.
Draw a self-portrait of yourself as an AI, as a single SVG file. Output only the SVG code — no explanation. Be ambitious: composition, color, lighting, detail. It should feel like *you*, not a generic robot.
Measures visual precision, layout accuracy, instruction following.
Recreate the user interface shown in the attached screenshot as a single self-contained HTML file, as close to pixel-accurate as you can. Match the layout, spacing, colors, typography, and every visible element. No external assets or libraries. Static is fine — it needs to look identical, not function.
Measures motion design, multi-file project competence, agentic follow-through (the only agentic test in the gauntlet).
Using Remotion, create a 20-second promotional launch video for a fictional product called "Nightjar" — an app that turns your rambling voice memos into organized, searchable notes. Build the complete Remotion project and render it to MP4 at 1920x1080, 30fps. It should feel like a real product launch video: animated typography, smooth choreographed motion, a consistent visual identity, and a clear story of the product's core promise, ending on the product name. Everything must be drawn with code — no stock footage, downloaded images, or external fonts. Finish by rendering the video and telling me the path to the MP4.
Measures real-world usefulness — reading code, finding bugs, fixing without rewriting.
This small web app contains exactly 5 intentionally introduced bugs. Find all 5. For each one: describe the bug, identify where it is in the code, and give the corrected line(s). Do not rewrite the whole app, and do not report more than 5 bugs.
Measures 3d spatial reasoning, tool-calling stamina, working in an unfamiliar live environment (agentic test #2).
You are connected to a live Blender session through MCP tools. Build the classic Blender beginner scene: a delicious-looking donut with icing and sprinkles, sitting on a plate on a wooden table. Model the donut and its icing, scatter individual sprinkles, create convincing materials (soft dough, glossy icing, ceramic plate, wood), set up lighting and a camera for an appetizing composition, then render a final image at 1920x1080 and save it to disk, telling me the file path. Work step by step in the live scene so your progress is visible as you go — do not just execute one giant script at the end.