GPT-6 Astra Decodes 1941 German Army Enigma Radio Message Unsolved for 83 Years in Ten Hours
Bloomberg product development coach Carter Leffen stated that using OpenAI's GPT-6 Astra Extra High, a 1941 German Army Enigma radio message dated July 10, consisting of 82 characters and archived under number MVUEH, was decoded in about 10 hours after remaining unsolved for 83 years.
The message was sent by a soldier stationed in Rosenow, requesting an immediate radio reply and providing a marching route. The German text includes "Sofort Funkantwort" and "Angabe des Marschweges," translated as "please specify the marching route, I am in Rosenow, please reply immediately;" the text also retains common misspellings by the operator, such as BTTE instead of BITTE.
The model is described as autonomously retrieving historical archives, comparing questionable letters, searching for contextual clues, building an Enigma simulator, writing code for cryptanalysis, conducting parallel experiments, testing competing keys, and cross-verifying. By utilizing the Enigma's property that "letters never encrypt to themselves" to eliminate positions and using place names from already decoded messages on the same day as cribs, the machine settings were calculated after a match was established, resulting in 68 remaining characters forming coherent German. A separate preserved header was used for verification of the settings.
This encrypted message comes from the 1941 German Army interception archives compiled by Frode Weierud and others, belonging to the short messages of the Northern Army Group's communications during the early stages of Operation Barbarossa that had not yet been broken. Some date keys from the same batch of materials had previously been manually restored; Leffen noted that there seem to be about 7 similar unsolved messages remaining in the 1941 archives.
Reports indicate that this operation consumed approximately 650 million tokens. Leffen mentioned that the effort he put into the demonstration website was about 99 times that of the decryption itself. OpenAI positions Astra as a model for end-to-end processing of high-difficulty tasks and has marked it as Critical within the cybersecurity readiness framework.
In market mechanisms, this is a capability demonstration-driven pricing, not a transaction of war history itself: what is bought is an agent's labor that can "write attack code, build simulators, and string archives"; what is sold relies on expert teams advancing classical cryptography and archival organization by the year. Beneficiaries include model vendors and cloud inference credits that need to prove end-to-end network/cryptographic capabilities to enterprise and defense clients; those under pressure are small cryptography history projects that treat unsolved historical encrypted messages as long-term topics, as well as traditional analysis outsourcing still priced by man-hours.
Additionally, the same model was also used for a trial decryption of another 1918 German radio message, cross-referencing locations and dates with naval logs; the Enigma itself was mathematically opened by Marian Rejewski back in 1932, and during World War II, the Allies relied on bombers and process scaling; this time, the increment is integrating retrieval, modeling, attack code, and verification into the same unattended link.
Source: Public Information
ABAB AI Insight
OpenAI has made Astra operable via browsers, capable of writing code and running network exploits, with the first batch of publicly showcased high-difficulty cases not being new papers, but rather using previously disclosed yet unresolved historical encrypted messages as benchmarks. The approach follows the same logic as previous FrontierMath, ExploitBench, and browser zero-day exploit assessments: using "problems that human experts have long left unresolved" as capability advertising. Leffen from Bloomberg provides the objectives and decisions, while the model handles archives, simulators, and key searches, closely resembling the division of labor in a laboratory where a principal investigator sets the problem and graduate students work overnight on attack scripts.
Capital has not flowed to museums but rather to inference volumes and agent orchestration. The 650 million tokens, Extra High tier, along with the demonstration site, indicate that the marginal cost of a single "archaeological-level decryption" has dropped to a level affordable for individual developers, but scaling to tackle the remaining 7 short messages from 1941, or shifting to modern protocols, requires ongoing computing power contracts rather than one-off news. The strategic motivation is to prove that Astra can generate its own toolchain on closed problems without existing solutions, which is more valuable for cybersecurity Critical ratings and enterprise orders on Azure/Bedrock than restoring a line of marching orders.
The analogy is not Turing's bombe hall, but rather the 2003–2005 work by Sullivan and Weierud on the Northern Army Group's encrypted messages, performing unique ciphertext attacks and leaving the remaining short messages in CryptoCellar for others to claim. The industry position has shifted from "model answering questions" to "model self-building laboratories": first solving classical cryptography, which has low verification costs and strong narratives, and then connecting the same agent stack to modern software exploits. Zodiac remnants and lists of unsolved classical ciphers will become the next batch of free benchmarks.
Structurally, this represents a technological substitution. In the past, breaking the Enigma required the division of labor among mathematicians, linguists, and electromechanical search machines; now, retrieval, cribs, simulators, and parallel key trials are integrated into the same model loop. The mechanism at play is that while the state space of classical cryptography is large, the rules are fully known and plaintext is highly formatted, making it perfectly suited for an agent that "writes code + runs experiments + prunes using historical context," without needing new mathematical breakthroughs. The unresolved issue is no longer the principles, but who connects the archives and computing power first.
ABAB News · Cognitive Laws
For known-rule problems, the outcome belongs to agents rather than expert compilers.
Unresolved archives are the cheapest capability advertisements.
When decryption costs drop to individuals, pricing power shifts to inference usage.