ExplorationBench - AlienCode (Open-Book): leaderboard
An open-book control of the AlienCode sandbox of ExplorationBench (Fudan University, Tencent Hunyuan and Tsinghua University, 2026): the complete set of altered language rules is supplied before any exploration, so the task measures applying an unfamiliar, fully specified program semantics. Percent of 70 held-out program-synthesis tasks solved, each answered three times in a single run; higher is better.
Metric: Held-out accuracy with the complete rule set supplied before exploration (%; open-book control, one run per system, no tool calls; AlienCode, 70 program-synthesis and prediction tasks graded by an interpreter on private inputs; 70 held-out tasks each answered three times in a tool-disabled copy of the conversation and averaged; highest reasoning setting each API offers). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Max) | 84.3 | #20 (Claude Opus 5) |
| 2 | Gemini 3.8 Flash (High) | 78.1 | #36 (Gemini 3.8 Flash) |
| 3 | GPT-5.6 Sol (Max) | 68.1 | #29 (GPT-5.6 Sol) |
| 4 | Kimi K3 (Max) | 52.4 | #53 (Kimi K3) |
| 5 | Grok 4.6 (xHigh) | 51 | #44 (Grok 4.6) |
| 6 | DeepSeek V4.1 Flash (Max) | 39.5 | #58 (DeepSeek V4.1 Flash) |
| 7 | DeepSeek V4 Pro (Max) | 37.6 | #181 (DeepSeek V4 Pro) |
| 8 | Hy4 preview (High) | 27.1 | #66 (Hy4 preview) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=explorationbench-aliencode-open-book · How It Works · Data refreshed daily, snapshot 2026-11-12.