01
No screenshot
pass
Anthropic
Claude Opus 4
Duration
22m 19s
Cost
—
Code lines
1,820
Files
11
Debug loops
0
Debug %
15%
Time breakdown
Initial Debug
RoundApril 16, 2026Combat Arcade Racer
April 2026 — Coding Tool Benchmark
AI coding model benchmark — Claude, Qwen, GLM, MiniMax, Kimi, and MiMo compared on a real-world Three.js game build.
8 coding models built the same Three.js game brief. Identical prompt, identical generated assets, identical agentic workflow. The only variable is the model.
Models
08
Pass rate
75%
Fastest
22m19s
Claude Opus 4
Published April 16, 2026Game Combat Arcade RacerGenre Combat Racing / Arcade Sim
Methodology
One variable: the coding model.
Prompt, assets, workflow, and strategy advice are identical across runs. Only the model changes.
Brief
Combat Arcade Racer
A high-octane street racer that merges responsive arcade-style car mechanics with aggressive power-up systems in dense metropolitan environments. Players navigate claustrophobic city circuits, utilizing tactical abilities to outmaneuver traffic and complete high-risk challenges in a quest for urban dominance.
Combat RacingArcade SimUrban SportsCompetitive
Held constant
- Concept & prompt
- One game design doc, shared verbatim.
- Generated assets
- Same concept art, 3D models, and audio.
- Workflow
- build_mid_no_strategy — agentic coder loop.
- Strategy advisor
- Pre-baked recommendations injected identically.
- Debug iterations
- Bounded loop. Model decides when to stop.
- Timeout
- 60 minutes per run. After that, it fails.
Leaderboard
Wall-clock time to a playable build.
Winner
Pass
Fail
Figure 1. Total elapsed time per model, sorted fastest-first. Failed runs pinned to the end.
Data table
Full per-model data.
| 01 | Claude Opus 4 | Anthropic | Pass | 22m 19s | — | 1,820 | 11 | 0 |
| 02 | Claude Sonnet 4 | Anthropic | Pass | 36m 47s | — | 9,225 | 14 | 4 |
| 03 | GLM-5.1 | Zhipu AI | Pass | 48m 10s | — | 2,150 | 13 | 5 |
| 04 | MiniMax M2.7 | MiniMax | Pass | 57m 0s | — | 1,980 | 12 | 6 |
| 07 | Qwen 3.5 35B | Alibaba | Fail | 60m 0s | — | 850 | 7 | 8 |
| 08 | Kimi K2.6 | Moonshot | Fail | 60m 0s | — | 620 | 5 | 8 |
| 05 | MiMo V2 Pro | Xiaomi | Pass | 66m 0s | — | 1,740 | 10 | 7 |
| 06 | Qwen 3.6 Plus | Alibaba | Pass | 79m 15s | — | 2,610 | 15 | 4 |
Tap any column to sort. Cost column fills in when pipeline telemetry lands.
Games
Output per model.
Per-model breakdown
Every model, per-run data.
02
No screenshot
pass
Anthropic
Claude Sonnet 4
Duration
36m 47s
Cost
—
Code lines
9,225
Files
14
Debug loops
4
Debug %
51%
Time breakdown
Initial Debug
03
No screenshot
pass
Zhipu AI
GLM-5.1
Duration
48m 10s
Cost
—
Code lines
2,150
Files
13
Debug loops
5
Debug %
43%
Time breakdown
Initial Debug
04
No screenshot
pass
MiniMax
MiniMax M2.7
Duration
57m 0s
Cost
—
Code lines
1,980
Files
12
Debug loops
6
Debug %
46%
Time breakdown
Initial Debug
05
No screenshot
pass
Xiaomi
MiMo V2 Pro
Duration
66m 0s
Cost
—
Code lines
1,740
Files
10
Debug loops
7
Debug %
47%
Time breakdown
Initial Debug
06
No screenshot
pass
Alibaba
Qwen 3.6 Plus
Duration
79m 15s
Cost
—
Code lines
2,610
Files
15
Debug loops
4
Debug %
11%
Time breakdown
Initial Debug
07
No screenshot
fail
Alibaba
Qwen 3.5 35B
Duration
60m 0s
Cost
—
Code lines
850
Files
7
Debug loops
8
Debug %
60%
Time breakdown
Initial Debug
08
No screenshot
fail
Moonshot
Kimi K2.6
Duration
60m 0s
Cost
—
Code lines
620
Files
5
Debug loops
8
Debug %
64%
Time breakdown
Initial Debug
Head-to-heads
Head-to-head comparisons.
Head-to-head
claude opus 4 vs claude sonnet 4
Claude Opus 4 and Claude Sonnet 4 produced playable builds from the same Three.js brief with a 5:1 gap in code volume.
Head-to-head
claude opus 4 vs glm 5 1
Claude Opus 4 finished the Three.js brief in 22 min 19 s; GLM-5.1 finished in 48 min 10 s.
Head-to-head
qwen 3 6 plus vs qwen 3 5 35b
Qwen 3.6 Plus passed in 79 min 15 s. Qwen 3.5 35B hit the 60-minute timeout.
Build your own
Build a playable game with Sandscape.
Sandscape takes a text prompt and returns a playable game. The same models in this benchmark run under the hood. No coding required.