Skip to content
GlanceHall
  • Builds
  • Models
  • Compare
  • How we test
Get the next round
  • Builds
  • Models
  • Compare
  • How we test
  • Journal
  • About
Get the next round

We don't rate. We let you see.

contact@glancehall.com

Watch

  • All builds
  • Models
  • Head to head
  • Journal

Trust

  • How we test
  • About
  • FAQ
  • Contact

Legal

  • Privacy
  • Terms
  • Cookies
© 2026 GlanceHallIndependent. Not affiliated with Anthropic, OpenAI or any model vendor.
Builds/Simulation

Balls in a spinning hexagon

Twenty numbered balls under gravity inside a hexagon that turns once every ten seconds. Physics written from scratch: moving-wall collisions, ball-to-ball contact, friction and visible spin. No libraries.

Simulation4 models · one shot eachRecorded Oct 22, 2026

Claude Opus 5.5Max effort
Built in 40m 11sErrors 0Open live
Claude Fable 5.1Max effort
Built in 25m 17sErrors 0Open live
GPT-6 AstraExtra high reasoning
Built in 8m 14sErrors 0Open live
GPT-6.1 SolExtra high reasoning
Built in 7m 54sErrors 0Open live

The facts

Measured, not judged. Nothing here is a score.

MeasureClaude Opus 5.5Claude Fable 5.1GPT-6 AstraGPT-6.1 Sol
Time to build40m 11s25m 17s8m 14s7m 54s
AttemptsOnly an infrastructure failure (a rate limit or an outage) earns a retry. Every interrupted attempt is listed, and nothing from it is published.1111
Delivered source31 KB15 KB24 KB25 KB
Files1111
Tokens in / outAs reported by each CLI. Tokenizers differ between vendors, so compare within a vendor.11.5M / 211k2.5M / 106k584k / 24k1.1M / 43k
Errors during recording0000
Video lengthUI recordings scroll at a fixed speed, so a longer page makes a longer video.20s20s20s20s

What to look for

Questions, not answers. The call is yours.

  1. Do any balls tunnel through a wall or escape?
  2. Does the rotating wall push the balls in a believable way?
  3. Do the numbers spin with the balls?
  4. Does the pile settle, jitter or sink into itself?

How it was recorded

No input: the first 20 seconds after page load, recorded frame by frame.

The same script runs on every build. Open any build live to judge sound, feel and anything a video can't show. Full method

The exact prompt

Sent word for word to every model. Try it on your own favourite.

Build a physics simulation in the browser: 20 balls bouncing inside a spinning hexagon.

REQUIREMENTS
- A regular hexagon, centred in the viewport and as large as comfortably fits, rotating continuously at a constant speed of one full turn every 10 seconds.
- 20 balls of equal size, numbered 1 to 20, each with a distinct colour. They start near the centre and fall under gravity.
- Implement the physics yourself, with no libraries: gravity, air damping, restitution when bouncing, ball-to-ball collisions, and collisions with the moving walls. Because the walls rotate, the velocity of the wall at the point of contact must affect each bounce.
- Friction between balls and walls makes the balls spin. Show each ball's rotation by drawing its number rotating with it.
- Balls must never escape the hexagon or pass through a wall, and must not sink into each other when they pile up.
- Use time-based updates (based on the elapsed time between frames) so the simulation behaves the same at any frame rate. Render with canvas 2D, driven by requestAnimationFrame.
- Fill the window and stay correct when the window is resized. It will be viewed at 1440x900 on desktop and 390x844 on a phone.
- Start automatically when the page loads; no interaction is required. The visual style is up to you.

DELIVERY RULES
- Work autonomously. Nobody will answer questions, so make reasonable assumptions and keep going until the simulation is finished.
- Deliver a single file, ./index.html, with all HTML, CSS and JavaScript inline. No build step, no libraries, no network requests of any kind.
- The page must work when index.html is opened directly in a current version of Chrome, and when it is served over HTTP.

Read more about this brief

Testing & methodAI coding benchmarks explained: SWE-bench, Terminal-Bench, LiveCodeBench, arenas, and what they can't showWhat the major AI coding benchmarks actually measure, how to read their scores, where vendors' numbers and public leaderboards disagree, and the things no coding benchmark can tell you.October 22, 2026Best AI for…Best LLM for coding: Opus 5.5 vs Fable 5.1 vs GPT-6 Astra vs GPT-6.1 Sol, same promptsWhich LLM is best for coding? Four frontier models, identical prompts, one attempt each. Measured time, tokens, cost per build and runtime errors, plus the builds themselves to judge.October 22, 2026Can AI build it?Vibe coding games: three games, four models, one prompt eachIs it possible to vibe code a game? We gave four AI models three game prompts, one attempt each, no edits: a physics toy, an FPS and a MOBA. Here is what you get, and how to prompt for it.October 22, 2026Coding agentsCodex projects: 16 finished builds, from landing page to MOBASixteen projects built by OpenAI's Codex with GPT-6 Astra and GPT-6.1 Sol, one prompt each, no edits: websites, a dashboard, a phone app, a kanban board, a physics toy and two games.October 22, 2026
PreviousAnalytics dashboardNextFirst-person shooter