AI Builder Eval

Game Bench

Can an AI build a playable browser game in one pass? Each model gets the same frozen prompt, a clean workspace, and no human feedback.

A personal project by Luiz Piccini, Head of Product at Tela.

Play without seeing the models Read the protocol
Runs scored 28

Playable builds scored with the four-pillar rubric.

Highest artifact score 95.5

GPT-6 Astra at $4.95.

Best under $1 91.2

GPT 6.1 Sol at $0.75.

Human feedback 0

No corrections or steering during a run.

Efficiency

Game-Bench Index vs Cost

Compare benchmark score against run cost in USD. All 28 ranked games appear here.

Higher is better on the Y-axis. The visible score scale starts at 30 to make model differences easier to read; cost uses a log scale. The dashed line is the Pareto frontier: nothing on the chart is both cheaper and better than a run on it, so it traces the best value at each budget.

Leaderboard

Click any column header to sort. Scores are out of 100; final score is weighted.

Play
GPT-6 Astra high 24/Sep/26 16m 6s 1.84M $4.95 92.6 99.8 96.2 98.5 95.5
Claude Fable 5.0 high 09/Jun/26 17m 32s 1.96M $6.73 91.5 98.0 95.2 96.8 94.2
Claude Sonnet 5.5 high 13/Oct/26 14m 54s 1.92M $1.34 92.0 98.0 88.4 88.0 91.4
GPT 6.1 Sol high 13/Oct/26 26m 11s 1.73M $0.75 90.7 98.0 90.2 87.5 91.2
Gemini 3.8 Flash high 06/Sep/26 13m 24s 9.18M $1.39 86.6 100.0 89.0 88.3 89.5
DeepSeek V4 Pro 0813 high 04/Sep/26 22m 37s 2.99M $0.23 86.2 96.0 86.6 87.8 88.0
GLM 5.3 high 05/Sep/26 4h 18m 28s 3.75M $2.57 83.4 92.0 87.6 85.3 86.0
Claude Opus 4.8 xhigh 02/Jun/26 24m 0s 6.3M $7.03 78.6 94.0 90.4 88.7 85.4
GPT-6 Sol high 23/Sep/26 12m 26s 3.11M $1.16 79.5 96.0 79.6 83.8 82.6
Claude Fable 5.1 high 24/Sep/26 27m 17s 2.67M $5.99 75.8 100.0 84.2 81.3 82.4
Claude Opus 5.5 high 23/Sep/26 24m 53s 3.23M $3.85 74.4 96.0 87.4 81.3 81.9
GPT-6 Luna high 23/Sep/26 23m 7s 4.10M $0.09 77.1 96.0 81.2 81.3 81.6
Claude Haiku 5.5 high 13/Oct/26 24m 13s 2.18M $0.31 80.5 92.0 80.2 73.0 81.0
GLM 5.2 high 22/Jun/26 40m 29s 507K $1.58 80.0 96.0 68.8 85.0 80.3
GPT-5 Codex xhigh 02/Jun/26 42m 0s 3.1M $4.03 73.3 90.0 76.8 74.2 76.8
MiMo v2.6 Flash high 23/Sep/26 1h 5m 3s 4.38M $0.10 70.5 92.0 76.0 80.0 76.5
GLM 5.1 high 22/Jun/26 28m 30s 578K $3.35 72.9 92.0 68.4 85.0 76.5
MiMo v2.6 Pro high 24/Sep/26 34m 2s 2.69M $0.20 69.1 96.0 76.0 72.3 75.3
Qwen3.8 Max 0902 high 24/Sep/26 1h 52m 30s 11.85M $4.91 67.8 92.0 81.0 69.0 74.9
MiniMax M3 high 01/Jun/26 34m 38s 387K $0.59 65.8 80.0 66.6 76.1 69.7
Kimi K2.6 high 02/Jun/26 26m 0s 317K $0.34 66.1 74.0 61.3 75.3 67.5
Qwen 3.7 Plus high 14/Jun/26 9m 22s 271K $0.08 51.5 80.0 68.0 45.0 58.9
MiMo V2.5 Pro max 01/Jun/26 6m 54s 754K $0.05 50.0 72.0 45.3 57.3 53.2
MiMo V2.5 Pro high 29/May/26 13m 27s 1.93M $0.11 43.9 74.0 49.6 65.8 53.1
MiMo V2.5 Pro medium 29/May/26 12m 0s 1.69M $0.09 48.0 74.0 50.3 40.7 51.4
Kimi K2.7 Code high 12/Jun/26 21m 32s 3.86M $0.95 42.5 68.0 58.4 45.0 50.7
Gemini 3.5 Flash high 02/Jun/26 17m 0s 2.0M $1.47 36.5 44.0 47.1 80.1 46.8
DeepSeek V4 Pro high 22/Jun/26 23m 6s 570K $0.10 36.9 48.0 30.4 55.0 39.7

Protocol

How the eval works

01

Frozen prompt

Every model receives the same game brief, same constraints, and same anti-drift guardrails.

02

Clean workspace

No memory, previous builds, project context, or hidden examples. The run starts from zero.

03

One shot

The model plans, builds, tests, and delivers without iterative human steering.

04

Playable result

The output is judged on whether it loads, plays, communicates its rules, and feels worth trying.

Score method

Public scores expose four use-case pillars. Gameplay blends design, fun, and onboarding. Adherence is prompt-following fidelity. Engineering combines cross-platform support, implementation quality, and QA. Art combines visual art, UI/UX, and audio. Final score = Gameplay 45, Adherence 15, Engineering 25, Art 15.

What this measures

Builder capability under a real artifact constraint

Game Bench asks models for a complete game: input handling, state, rendering, tutorial clarity, play feel, failure states, and enough polish that a human can actually evaluate it. The idea is to test how models working at tasks closer to real life usage, not just coding exercises.

Reproducibility

Full frozen prompt

Read the exact prompt every model received: same game brief, same constraints, same one-shot mandate, same mobile and playtest requirements.

Open frozen prompt

Next runs

Let me know what other models or harnesses you think should be added

Send model names, coding agents, IDE harnesses, or run conditions that would make the benchmark more useful to compare.

Suggest a model or harness →

Why I run this

I’m Head of Product at Tela, where we help large companies put AI to work in their core operations. This means interacting and testing with many models and understanding which is the best at what job.

Game Bench is my personal project to test models in public in a fun way that is dear to me: what are the best models for making games?

Explore Tela ↗