Some ciphers surrender in hours. Others hold out for a century. This week, a World War I German radio message encrypted with the ADFGVX method finally gave up its secret — not to a human cryptographer, but to OpenAI's GPT-6 Astra.
The breakthrough, documented by cryptology researcher "prinz," targeted one of 50 unsolved ciphers listed on Scienceblogs.de, a German science blogging portal. The list ranges from cryptograms published by serial killers to the famous Voynich manuscript. Among them sat a set of German radio messages from WWI, transmitted in the final weeks of the war. One, sent on November 27, 1918 — more than two weeks after the Armistice — had resisted every expert effort, including work by veteran codebreaker George Lasry, who has personally decoded hundreds of these messages.
What Is ADFGVX, and Why Was It So Hard?
ADFGVX was Germany's most sophisticated field cipher of the Great War. Every letter (and digit) in the message is first converted into a pair of symbols drawn from just six characters — A, D, F, G, V, X — using a Polybius square built from a secret keyword. The resulting symbol pairs are then scrambled by a second, columnar transposition keyed to the same secret word. Breaking it requires solving two linked puzzles at once: the substitution layer and the transposition layer, with no way to attack one without guessing the other.
Hundreds of German ADFGVX messages have already been decoded, largely thanks to modern hill-climbing algorithms and the painstaking manual work of experts. But a stubborn residue of messages has resisted everything — the one GPT-6 Astra just cracked among them.
How the Model Did It
What makes this result remarkable isn't just the solve — it's the method. GPT-6 Astra didn't brute-force the message. It reconstructed the actual encryption key: TRUPPENVERSCHIEBUNG ("troop relocation"), a term documented on pages 214–215 of J. Rives Childs's The History and Principles of German Military Ciphers, 1914–1918, a period-accurate source.
From there, the model executed the full decryption pipeline the way a human analyst would:
- Alphabetically rearrange the 19-letter keyword to derive the column permutation
- Write the 170 transposed symbols under the keyword in rows of 19
- Read out the columns in permuted order to recover the substitution layer
- Apply the ADFGVX Polybius square to yield plaintext
The recovered message reads: "EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X" — in English: "An English cruiser arrived at Sevastopol on the ?4th. An Allied squadron follows on the 26th." A century-old intelligence report about Black Sea naval movements, suddenly legible again.
Why This Matters Beyond the History Books
Three takeaways for anyone watching the AI space:
- Long-horizon reasoning is real. This wasn't a one-shot pattern match. The model had to hold a multi-stage cryptanalytic procedure in context, validate intermediate results against period documents, and iterate. That's agent-grade reasoning, not autocomplete.
- History and archival science get a new tool. Thousands of undeciphered historical texts — military, diplomatic, personal — exist in archives. Models capable of hypothesizing keys and testing them cheaply could unlock a generation of historiography.
- The frontier moved from "benchmark" to "first-ever." Benchmark scores measure relative performance. Solving a problem no human ever solved is a different category of result — and it's becoming more common on frontier models.
Try the Model That Did It
You don't need an OpenAI enterprise contract to run frontier models on real problems. On Qubax's model marketplace, GPT-6 Astra is available at a fraction of retail pricing — roughly $1.05 per million input tokens and $4.21 per million output tokens, versus OpenAI's retail $5/$25. That's about 79% off, because Qubax is an open market where compute providers compete on price.
Whether you're cracking ciphers, analyzing archives, or building the next research agent, the full model catalog and API docs will get you started in minutes.
FAQ
Did GPT-6 Astra really solve the cipher unaided?
The researcher provided the ciphertext and pointed the model at relevant historical literature. The key reconstruction, the decryption procedure, and the final plaintext were produced by the model. As with most AI-assisted research, a human verified the result.
What is ADFGVX encryption?
A two-stage German field cipher from WWI: a Polybius-square substitution (converting each letter to an ADFGVX symbol pair) followed by a keyword-based columnar transposition. France broke it in 1918, but individual messages with unknown keys can still resist modern solvers.
How much does GPT-6 Astra cost?
Retail pricing is $5 per million input tokens and $25 per million output tokens. On Qubax, it currently runs at about $1.05 in / $4.21 out — roughly 79% below retail.
Can AI models break modern encryption?
No. Ciphers like ADFGVX are breakable because they're nearly 110 years old and historically weak by modern standards. AES-256 and current public-key cryptography remain computationally infeasible to break with any existing model or hardware.
