# ExplorationBench AlienCode Best@3 (ExplorationBench AlienCode)

> Tests whether a system can discover a programming language's hidden rules through experiments and apply them to unseen tasks.

Canonical page: https://benchlm.ai/benchmarks/explorationbench-aliencode

- Category: [Agentic](/agentic)
- Last updated: October 15, 2026 snapshot

## About ExplorationBench AlienCode

- Year: 2026
- Tasks: 70 held-out program-synthesis tasks; 31 discovery targets
- Format: Held-out accuracy (Best@3)
- Difficulty: Discovering and applying unfamiliar executable rules
- Paper: [ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds](https://arxiv.org/abs/2609.30199v1)

Four exploration rounds precede tool-disabled evaluation on 70 tasks. The interpreter checks each program on five private inputs. The headline is the best complete run of three; the table also shows the mean and worst. Scores follow the benchmark-owned web table, with private code and tasks, and stay outside overall and category rankings.

ExplorationBench AlienCode is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (10 systems)

| Rank | System | Reasoning | Before exploration | Best@3 | Mean@3 | Worst |
|---|---|---|---|---|---|---|
| 1 | [Claude Opus 5](/models/claude-opus-5) | max | 3.8% | 89.0% | 73.5% | 49.0% |
| 2 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | max | 11.0% | 88.1% | 79.5% | 73.8% |
| 3 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | high | 1.4% | 79.0% | 39.7% | 6.2% |
| 4 | [Kimi K3](https://explorationbench.com/) | max | 2.4% | 79.0% | 38.7% | 5.7% |
| 5 | [Grok 4.6](/models/grok-4-6) | xhigh | 4.8% | 69.0% | 53.2% | 37.1% |
| 6 | [Qwen3.8-Max](/models/qwen3-8-max) | max | 2.4% | 65.7% | 64.0% | 61.4% |
| 7 | [Hy4 preview](/models/hy4-preview) | high | 0.5% | 63.3% | 30.2% | 1.4% |
| 8 | [DeepSeek-V4.1-Flash](/models/deepseek-v4-1-flash) | max | 1.4% | 59.5% | 46.3% | 23.3% |
| 9 | [Seed2.1 Pro](/models/seed-2-1-pro) | high | 2.4% | 40.0% | 24.8% | 4.8% |
| 10 | [DeepSeek-V4-Pro](https://explorationbench.com/) | max | 1.9% | 12.9% | 7.3% | 2.4% |

Before exploration and Best@3 come from the same selected run. Mean@3 and Worst cover all three runs; each checkpoint averages three answers per task.

Scores follow the current web table. The September 24 paper describes the protocol but reports an earlier AlienCode maximum of 87.6% versus 89.0% on the website. Code and held-out tasks remain private.

Sources: [Official leaderboard](https://explorationbench.com/), [published result data](https://explorationbench.com/assets/js/results.js), and [paper and protocol](https://arxiv.org/abs/2609.30199v1).

## FAQ

### What does ExplorationBench AlienCode measure?

Tests whether a system can discover a programming language's hidden rules through experiments and apply them to unseen tasks.

### Which model leads the published ExplorationBench AlienCode snapshot?

Claude Opus 5 currently leads the published ExplorationBench AlienCode snapshot with a score of 89.0%.

### How many models are evaluated on ExplorationBench AlienCode?

The October 15, 2026 snapshot contains 10 systems at their published reasoning settings.

### Does ExplorationBench AlienCode affect BenchLM's overall score?

Not directly. ExplorationBench AlienCode is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.