OpenAI's Astra and Anthropic's Claude Opus 5 have successfully decoded two Enigma messages that had resisted human cryptanalysts for decades, including one unbroken since 2005. The breakthrough signals a shift in how frontier language models (LLMs) can tackle complex research problems, moving beyond simple pattern recognition to perform the full workflow of a research analyst.

What Exactly Did These AI Models Accomplish?

Developer Carter Leffen used OpenAI's Astra to search a database of Enigma messages, identify an unbroken one, and decode it without significant human guidance. Astra performed archival research, gathered context clues, built its own simulator of the Enigma machine, and recovered the plaintext of the 2005 message in just two days. The validation came from Frode Weirerud, a retired electrical engineer who maintains Crypto Cellar, a long-running repository of Enigma records. Weirerud confirmed Leffen's solution and said it left him in "awe".

On September 21, cryptanalyst Jack Willis used Anthropic's Claude Opus 5 to crack a second unsolved message, though with more hands-on human guidance. Claude Opus 5 exploited the known signature of a particular officer's name to work back to the plaintext. These two breaks leave just seven unbroken Enigma messages in the known archival record, plus one where the plaintext is known but the code itself remains unbroken.

How Do These AI Breakthroughs Differ From Traditional Codebreaking?

The original Enigma break, led by Alan Turing at Bletchley Park during World War II, required the construction of the Bombe, an early electromechanical computer purpose-built to search Enigma's key space. That machine brute-forced a known cipher structure through raw computational power.

What Astra and Claude Opus 5 accomplished is fundamentally different. Rather than simply brute-forcing a cipher, these models handled the complete workflow a human cryptanalyst would perform: locating candidate messages in archives, reasoning about likely context and senders, building the tooling to test hypotheses, and iterating to a solution. This is closer to the job of a research analyst than to the job of a traditional codebreaker.

"Weirerud, who has spent decades on this material, said he personally spent several weeks working through the same Bundesarchiv files that Astra parsed in two days," noted the analysis of this breakthrough.

Frode Weirerud, Crypto Cellar Repository Maintainer

What Are the Key Differences Between the Two Approaches?

• Astra's Method: Performed independent discovery with minimal human scaffolding, conducting archival research, building its own Enigma simulator, and reaching a solution autonomously over two days.
• Claude Opus 5's Method: Required significant human guidance, with cryptanalyst Jack Willis supplying the officer's-name signature that unlocked the message, making its role closer to assisted decryption than independent discovery.
• Archival Access: Weirerud noted uncertainty about how Astra accessed archived messages held in a "private collection" not hosted on Crypto Cellar, speculating that another researcher may have posted them online or that Astra pulled from the German government's public Bundesarchiv holdings.

What Does This Mean for Future AI Research Applications?

The broader signal from these breakthroughs is what frontier LLMs can now accomplish when pointed at narrow, verifiable problems with a clear success condition. Cryptanalysis of historical ciphers serves as an ideal testbed: the answer is checkable, the corpus is bounded, and the skill ceiling is high.

Experts expect more of the remaining seven unbroken Enigma messages to fall to similar AI-assisted approaches. The same pattern is likely to migrate into live scientific problems where the checkable-answer property still holds. This represents a new category of AI application: the model as tireless research analyst, with humans serving as validators rather than primary solvers.

How to Evaluate AI Breakthroughs in Research

• Reproducibility Matters: The provenance of archival material matters for reproducibility, even if it does not affect the validity of the plaintext itself, so researchers should document exactly which sources the AI accessed.
• Human Scaffolding Varies: Some AI solutions require minimal human input while others need significant guidance, so the level of human involvement should be clearly disclosed when reporting results.
• Verification by Experts: Independent validation by domain experts with decades of experience is essential to confirm that AI-generated solutions are genuinely correct, not plausible-sounding errors.
• Speed as a Metric: The compression of research timelines from weeks to days represents a meaningful efficiency gain, but speed alone does not validate correctness without expert review.

These Enigma breakthroughs underscore a pivotal moment in AI capability. Frontier models like Astra and Claude Opus 5 are moving beyond answering questions to solving research problems that require reasoning, tool-building, and iterative hypothesis testing. The fact that these models can now tackle problems that stumped human experts for two decades suggests that the next wave of AI impact may come not from replacing human researchers, but from amplifying their capacity to tackle problems at scale.