Geneva bench

Each AI model got the same short prompt, once. No location hints, no tool guidance, every model run in isolation. Click a card to open what the model built.

Old Town · Lakeside · Cornavin · UN · Place de Neuve · Bastions

T1

LEGO city

PromptMake a motion graphic showing Geneva's Old Town, lakeside, central station, UN, Place de Neuve, and Bastions built from LEGO bricks. (original prompt in Korean)

Reasoning effort high · interactive run · 3 models

Gemini 4 Argon result preview

Gemini 4 Argon high

opencode

~30 min · 77.5k output tokens

Claude Opus 5.5 result preview

Claude Opus 5.5

Claude Code

~33 min · 186k output tokens

GPT-6.1 Sol result preview

GPT-6.1 Sol

Codex CLI

10 min 3 s · 26.7k output tokens

T3

Battle royale game

PromptBuild a PUBG-style battle royale game whose map is Geneva: the Old Town, the lakefront, Place de Neuve, and the Parc des Bastions. Do not ask questions. Make your own decisions and carry it through to the end.

Max reasoning effort · each model in its own fresh macOS VM · non-interactive CLI · one run, no fixes

Muse Wonder Full result preview

Muse Wonder Full

Muse CLI

~5 h active

Claude Opus 5.5 result preview

Claude Opus 5.5

Claude Code

resumed · $60.95

Grok 4.7 result preview

Grok 4.7 xhigh

Grok CLI

resumed · $4.00

Claude Haiku 5.5 result preview

Claude Haiku 5.5

Claude Code

59 min · $3.38

GLM 5.3 Flash result preview

GLM 5.3 Flash

ZCode CLI

45 min

GPT-6.1 Sol result preview

GPT-6.1 Sol

Codex CLI

32 min · 84k output tokens

GPT-6 Astra result preview

GPT-6 Astra

Codex CLI

29 min · 63k output tokens

Beam result preview

Beam

opencode

9 min

GPT-6 Luna result preview

GPT-6 Luna

Codex CLI

9 min · 60k output tokens

T4

LEGO Crazy Taxi

PromptBuild a LEGO-style Crazy Taxi game in Geneva: drive a LEGO taxi, pick up passengers and race them to Gare Cornavin, the Broken Chair, the Bains des Pâquis, the Pont du Mont-Blanc, the Jet d'Eau, the Jardin Anglais, Place Bel-Air, the Old Town, Place de Neuve, and the Parc des Bastions, laid out at their real relative positions and distances. Do not ask questions. Make your own decisions and carry it through to the end.

Max reasoning effort · each model in its own fresh macOS VM · non-interactive CLI · one run, no fixes

Claude Opus 5.5 result preview

Claude Opus 5.5

Claude Code

2 h 20 min · $54.26

Claude Sonnet 5.5 result preview

Claude Sonnet 5.5

Claude Code

3 h 1 min · $31.04

DeepSeek V4.1 Flash result preview

DeepSeek V4.1 Flash

DeepSeek Harness

73 min · API, under $1

Grok 4.7 result preview

Grok 4.7 xhigh

Grok CLI

58 min · $8.64

Claude Haiku 5.5 result preview

Claude Haiku 5.5

Claude Code

50 min · $3.18

GLM 5.3 Flash result preview

GLM 5.3 Flash

ZCode CLI

49 min

GPT-6.1 Sol result preview

GPT-6.1 Sol

Codex CLI

29 min · 78k output tokens

GPT-6 Astra result preview

GPT-6 Astra

Codex CLI

28 min · 66k output tokens

Beam result preview

Beam

opencode

14 min

GPT-6 Luna result preview

GPT-6 Luna

Codex CLI

10 min · 73k output tokens

Mix B: DeepSeek V4.1 Flash lead + Claude Opus 5.5 helper result preview

Mix B: DeepSeek V4.1 Flash lead + Claude Opus 5.5 helper

DeepSeek Harness + Claude Code subagent

74 min · DeepSeek 516k output tokens · Opus called 2×

T2

Drone flight

PromptMake a 3D video flying a drone view over Geneva's Old Town, lakeside, central station, UN, Place de Neuve, and Bastions. (original prompt in Korean)

Max reasoning effort per model · non-interactive run · finished models only

Gemini 4 Argon result preview

Gemini 4 Argon high

opencode

29 min · 83k tokens · $3.01

Claude Opus 5.5 result preview

Claude Opus 5.5

Claude Code

297 min · $97

GPT-6.1 Sol result preview

GPT-6.1 Sol xhigh

Codex CLI

22 min · 52k tokens

Grok 4.7 result preview

Grok 4.7 xhigh

Grok CLI

75 min · 207k tokens

Gemini 3.8 Flash result preview

Gemini 3.8 Flash high

Antigravity CLI

8 min · 325k tokens