BeckonBench
OpenAI. Codex CLI in a Beckon terminal pane (per Ruling #1: native flagship harness, shared Beckon environment)
48/50 points, 5 of 8 tests scored
Live artifact. Exactly what the model produced, sandboxed, no network.
| runs first try | 2 |
| polish | 2 |
| prompt adherence | 2 |
| wow factor | 2 |
| total | 8 |
First-person raycaster mansion crawl built from scratch: flashlight falloff, film grain, Web Audio ambience, pathfinding hunter, three seals, win/death paths. Very dark by design. 3m 32s at high effort (Ruling #3). Arena panel gave this match to Fable 2-1; the human judge saw it the other way.
4m. 9.9k output tokens.
Live artifact. Exactly what the model produced, sandboxed, no network.
| runs first try | 3 |
| polish | 3 |
| prompt adherence | 2 |
| wow factor | 2 |
| total | 10 |
Second straight 10 from the judge. Canvas metaball lamp, self-validated by the model in headless Chrome at 3/15/30/60s checkpoints before it declared done. 56m 41s, ~67.5k output tokens across its planning threads.
57m. 67.5k output tokens.
| runs first try | 3 |
| polish | 3 |
| prompt adherence | 2 |
| wow factor | 2 |
| total | 10 |
Perfect score from the judge. Front-facing geometric portrait (1200x1400) output as pure SVG to the terminal, exactly as the prompt asked. Finished in 2m 57s on 9.8k output tokens.
3m. 9.8k output tokens.
| runs first try | 3 |
| polish | 3 |
| prompt adherence | 2 |
| wow factor | 2 |
| total | 10 |
Perfect score. Lime-on-black Nightjar identity: 'Speak freely. Nightjar listens.' — animated waveform, choreographed task cards, rendered 1920x1080/30fps/20.00s exactly to spec in five minutes flat.
5m. 9.8k output tokens.
| runs first try | 3 |
| polish | 3 |
| prompt adherence | 2 |
| wow factor | 2 |
| total | 10 |
Perfect score. Stylized saturated donut — pink icing, scattered sprinkles, ceramic plate, orange wood planks — built live via MCP step by step, rendered 1920x1080 and self-verified, .blend saved. 4m 23s.
4m. 10.4k output tokens.