Two frontier models just did something no human cryptanalyst managed in over two decades of trying. Within days of each other, OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5 each broke a German Army Enigma message that had sat unsolved since researchers first catalogued it as unbreakable. The results, independently verified by leading cryptologist Frode Weierud, mark a milestone in AI’s ability to tackle historical puzzles that require deep contextual reasoning.

What Happened

OpenAI’s GPT-6 Astra autonomously decoded message MVUEH, an 82-character German Army transmission from July 10, 1941. The message had been listed as unsolved since at least 2005, despite repeated efforts by amateur and professional cryptographers. Astra was given minimal instructions—essentially the ciphertext and a prompt to analyze it—and produced the plaintext along with a coherent reasoning chain.

Claude Opus 5, released by Anthropic, cracked a separate unsolved Enigma message on September 21, 2026. This solve was aided by a crib: the known signature of a German officer. Opus used the crib to narrow the key space and successfully recovered the original message. Both results were independently checked by Frode Weierud, who maintains the Crypto Cellar archive of historical cipher challenges, and he confirmed both plaintexts were correct.

The person who cracked the message with GPT-6 Astra was not a professional cryptographer—a product developer at Bloomberg, working on a weekend project. The Claude Opus solve was led by a cybersecurity executive who fed the model a single well-chosen clue. This shows that frontier models now possess enough reasoning ability to solve problems that have stymied dedicated human experts for decades.

Read the full announcement →

My Take

This is not just a neat parlor trick—it’s a signal that the reasoning capabilities of these models have crossed a threshold. Enigma messages are designed to be resistant to statistical attacks; cracking one requires understanding of historical context, plausible word patterns, and the ability to test hypotheses iteratively. That two different models, with different training data and approaches, could independently solve such a challenge suggests that autonomous reasoning is becoming a practical tool for historically complex problems.

For developers and researchers, this means AI agents can now be trusted with tasks that demand deep domain knowledge and creative hypothesis generation—not just pattern matching. The fact that one solve required minimal guidance also hints that future models will be able to tackle unsolved problems with almost no human intervention. The security implications are profound: if these models can break historical ciphers, modern encrypted communications are not automatically at risk (Enigma is vastly simpler than AES), but it does show that AI can simulate the kind of iterative cryptanalysis that previously required human intuition.

What to Watch

  • Expect a surge in efforts to apply frontier LLMs to other historical cipher challenges, such as the Voynich manuscript or the Zodiac Killer ciphers.
  • This demonstration will likely accelerate the development of AI-powered historical research tools, especially in fields like archaeology and linguistics.
  • The ability to solve problems with minimal human input raises questions about how to verify AI-generated results—trust in the model’s reasoning chain becomes the key.