CipherBank: leaderboard
Decrypting privacy-sensitive texts enciphered with nine algorithms, from ROT and Vigenere to three invented ciphers. The model is not told the algorithm: it gets three plaintext and ciphertext example pairs and must infer the cipher before decrypting. Exact-match accuracy averaged over the nine ciphers. Higher is better. Source: arxiv.org/abs/2504.19093.
Metric: Decryption Accuracy (%; 3-shot, mean of nine ciphers). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 45.14 | #303 |
| 2 | O1 (2024-12-17) | 40.59 | #159 |
| 3 | DeepSeek R1 | 25.91 | #266 |
| 4 | O1 Mini (2024-09-12) | 20.07 | #414 |
| 5 | Gemini 2.0 Flash (Thinking) | 13.49 | #350 (Gemini 2.0 Flash) |
| 6 | DeepSeek V3 | 9.86 | #334 |
| 7 | Gemini 1.5 Pro | 9.54 | #453 |
| 8 | GPT-4o (2024-08-06) | 8.82 | #344 |
| 9 | Gemini 2.0 Flash (Preview) | 8.65 | #376 |
| 10 | GPT-4o (2024-11-20) | 6.4 | #392 |
| 11 | GPT-4o Mini (2024-07-18) | 1 | #566 |
| 12 | QwQ 32B-Preview | 0.76 | #672 |
| 13 | Qwen 2.5 72B Instruct | 0.55 | #470 |
| 14 | Llama 3.3 70B Instruct | 0.42 | #541 |
| 15 | Llama 3.1 70B Instruct | 0.38 | #570 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cipherbank · How It Works · Data refreshed daily, snapshot 2026-11-07.