GPT-6 Astra might have cracked a World War I radio cipher

Why it matters

Astra connected archival cryptography to independent naval records in a single research workflow. Without a reproducible transcript and expert review, the result remains a case study rather than a benchmark.

A writer publishing as prinz says OpenAI's GPT-6 Astra decoded a German radio message transmitted on November 27th, 1918, producing a plaintext whose central details match surviving Royal Navy records.

The September 17th post identifies the message as one of the remaining unsolved ciphertexts produced with ADFGVX, a German field cipher that combined substitution and columnar transposition. Astra's proposed reading says an English cruiser entered Sevastopol on the 24th and an Allied squadron would follow on the 26th.

Those dates line up with the movements of HMS Canterbury. The available historical account says HMS Canterbury arrived in Sevastopol on November 24th, while an Allied squadron followed on November 26th.

A British memorandum preserved by the Australian War Memorial supplies a second historical check. Dated November 24th, it says Canterbury and two other vessels had left for Sevastopol on November 23rd, while a larger multinational squadron was scheduled to depart Constantinople on November 25th and rendezvous the following morning.

That documentary match makes Astra's plaintext credible. It does not establish that this is the first correct solution, or that the model recovered it without decisive clues supplied in the archival material.

What Astra appears to have done

ADFGVX was introduced by the German Army in June 1918. The system converted letters and numbers into pairs drawn from six symbols: A, D, F, G, V and X. It then scrambled those symbols through a columnar transposition controlled by a separate keyword.

A 2016 Cryptologia paper by George Lasry, Ingo Niebel, Nils Kopal and Arno Wacker documented the surviving Eastern Front traffic. The researchers said James Rives Childs, an American signals-intelligence officer, had preserved 460 intercepted messages comprising 668 individual cryptograms. Their computerized methods decrypted 618 of them.

The candidate examined by Astra contains 170 symbols. According to prinz, the model applied the 19-letter transposition keyword "TRUPPENVERSCHIEBUNG," which appears on pages 214 and 215 of Childs' archival material. The target ciphertext appears on page 217.

With 170 symbols arranged beneath a 19-letter keyword, the transposition produces 18 columns containing nine symbols and one containing eight. Reordering those columns according to the alphabetic order of the keyword reconstructs the stream of ADFGVX coordinate pairs. A substitution table then converts each pair into plaintext.

The proximity of the keyword matters when assessing the result. Astra appears to have selected and applied a historically documented key, then tested the resulting text against external records. That is a useful research workflow involving archival reading, arithmetic, decryption and source checking. It is a key-application and validation task rather than a demonstrated ciphertext-only recovery of an unknown key.

The date discrepancy remains the technically interesting part. Prinz writes that the keyword was associated with messages beginning on December 9th, 1918, nearly two weeks after the November 27th transmission. Astra proposed that the unexpected reuse or early use of the key may explain why earlier attempts missed the message. The post does not establish whether the date assignment, intercepted message or archival annotation contains an error.

The naval records carry the result

The proposed German plaintext is imperfect. It contains a truncated form of the verb for entering or arriving and an uncertain character in the date. Its meaning is still clear enough to describe an English cruiser at Sevastopol on November 24th and a following Allied squadron on November 26th.

Both facts appear in contemporary naval material. The match is especially specific because it combines a vessel type, location and two dates separated by the same interval described in the proposed message. The original post accidentally gives Canterbury's arrival year as 2018 in one sentence; the linked log and the underlying events are dated 1918.

OpenAI released GPT-6 Astra on September 3rd and positioned it for complex research, science, browsing and long-running professional tasks. The cipher experiment offers a more concrete illustration than another benchmark percentage: the model connected an obscure archival procedure to records that could test the semantic output.

The experiment still lacks the materials required for a reproducible evaluation, including a complete interaction transcript and an independent cryptographic review. The defensible result is narrower than the headline claim circulating around it. Astra generated a historically coherent decrypt with a documented key, and separate naval records support its two central facts. The claim that Astra produced the first solution remains unverified.