A recent experiment tested nine AI models on a 54-character cryptographic puzzle involving a Rubik's Cube, originally proposed by Klaus Schmeh in 2018. The task required models to provide the plaintext, initial layout, and the sequence of moves used to scramble the cube within a 20-minute limit without internet access.
Only Claude (Opus 5.5) successfully solved the puzzle in approximately 6 minutes. The model identified the plaintext as 'MEET YOU AT THE WOODEN BRIDGE IN THE CITY TOMORROW EVENING' and provided a 10-move key. The solution was verified by a custom program, confirming it matches the ciphertext exactly.
Other models, including Gemini, ChatGPT (GPT-6 Astra Max), DeepSeek, Qwen, GigaChat, Alice, and Grok, failed to solve the puzzle independently. Gemini provided a false solution with a fabricated proof, while others either admitted failure or provided incomplete analysis despite correctly identifying some constraints.
The experiment highlighted that success on one cryptographic task does not guarantee performance on similar ones. For instance, while GPT-6 Astra previously solved a different puzzle by Andrey Maslov, it could not solve the Schmeh challenge on the first attempt, only succeeding after being provided with the solution method.
