TapTap Maker Benchmark
TapTap Maker / Coding · Writes working Lua features in a real game engine. Engine execution determines scores, without a model judge.
What it measures and how
Only official main_board L2 pass rate is used. Cost, speed, Held-out, and category scores receive no extra votes. A fixed tier priority selects the representative row without consulting scores.
How this evidence is used
Receives 3.6% of the coding-and-design budget, with engine and L2 metric fixed.
Evaluation results
Fixed rules select one representative configuration per public model. Anonymous test models are omitted. Original ranks are retained, with up to 30 models per evaluation.
| Source rank | Source model name | Raw result | Representative configuration |
|---|---|---|---|
| 1 | gpt-6-astraOpenAI | 85.8% | xHigh reasoning |
| 2 | claude-fable-5-1Anthropic | 85.5% | High reasoning |
| 3 | claude-opus-5-5Anthropic | 85.5% | xHigh reasoning |
| 4 | gpt-6.1-solopenai | 85.3% | xHigh reasoning |
| 5 | claude-sonnet-5-5anthropic | 83.6% | xHigh reasoning |
| 7 | gemini-3.8-flashGoogle | 79.8% | High reasoning |
| 8 | gpt-5.6-solOpenAI | 79.8% | xHigh reasoning |
| 9 | grok-4.7-xhighxai | 77.8% | xHigh reasoning |
| 10 | gpt-6-solOpenAI | 76.8% | xHigh reasoning |
| 11 | gpt-6-lunaOpenAI | 76.6% | xHigh reasoning |
| 12 | qwen3.8-max-0902Alibaba | 76.2% | xHigh reasoning |
| 13 | glm-5.3Z.ai | 76.2% | Max reasoning |
| 14 | deepseek-v4.1-flashDeepSeek | 76.1% | Max reasoning |
| 15 | qwen3.8-flashAlibaba | 76.0% | xHigh reasoning |
| 16 | claude-haiku-5-5— | 75.8% | xHigh reasoning |
| 17 | glm-5.3-flashZ.ai | 75.0% | Max reasoning |
| 18 | gpt-5.6-terraOpenAI | 74.2% | xHigh reasoning |
| 19 | hy4-previewTencent | 72.5% | High reasoning |
| 20 | kimi-k3Moonshot AI | 72.3% | Source default |
| 21 | gpt-5.6-lunaOpenAI | 72.2% | xHigh reasoning |
| 22 | doubao-seed-evolvingByteDance | 71.6% | High reasoning |
| 23 | deepseek-v4-proDeepSeek | 70.5% | Max reasoning |
| 25 | deepseek-v4-flashDeepSeek | 69.1% | Max reasoning |
| 26 | gemini-3.1-pro-previewGoogle | 59.5% | Source default |
| 27 | MiniMax-M3MiniMax | 54.1% | Source default |
| 28 | claude-haiku-4-5Anthropic | 38.5% | High reasoning |
Evaluation limits & attribution
One engine and language do not generalize to other stacks. Models may use different agent drivers; the closed harness prevents task-level reproduction.
Data license: Official results are public. Redisplay retains attribution; full redistribution remains subject to Yiwan/TapTap terms.
Results are published by TapTap Maker. Original scores and the NexGateHub consensus index use different scales and cannot be added directly.