GPT-6 Astra and Opus 5.5 crack old ciphers and trace alchemical texts
Writing on the resobscura Substack, a historian argues that this week's dueling releases of GPT-6 Sol and Opus 5.5 mark a turning point: frontier models can now be used not just as research assistants for transcription but to actually attempt to solve open historical problems. The piece walks through several cases built with GPT-6 Astra and Opus 5.5.
The most striking is a July 10, 1941 Enigma message that had long resisted decipherment. Historical cryptological researcher Frode Weierud reports that the breakthrough was not in the codebreaking math itself, but in GPT-6 Astra noticing an overlooked note Weierud had posted in July 2026 about newly found collections of radio messages in the German Bundesarchiv, including material from the SS-Totenkopf Division's logistics command. Weierud says his team is still analysing the model's logs and does not yet fully understand how it located everything it referenced, including a mention of a private collection whose source is unclear.
The author also set GPT-6 Astra on John Dee's coded manuscript Liber Loagaeth, transcribed with the help of the "scryer" Edward Kelley. The model's conclusion was that the supposed code is mostly nonsense syllables rather than real cipher text, though it did find that Kelley's transcriptions grew more repetitive after a certain date (cross-checked against Dee's diary) and that one passage does encode a real reference, to the angelic figure Bornogo. Separately, GPT-6 Astra identified the previously unidentified French alchemical source that Isaac Newton had freely translated into Latin. Attempts to find new, unsolved WW1 and WW2 era encrypted messages came up empty.
Opus 5.5 was set loose on the papers of Samuel Hartlib, the self-described "intelligencer" and early Royal Society figure, downloading more than 5,000 primary source files and deploying sub-agents to cross-check unidentified sources of Hartlib's information across languages via Google Books and other archives. Its preliminary finding: both Newton and Hartlib used their own anagrams for "Hungarian vitriol." Newton's coded term "Vltimorui" is a true anagram of vitriolum, already identified as such by the Newton edition's editors; Hartlib's "Miloirtiua riciragnun" is closer to a reversed spelling of the same word plus "ungaricum." Combined with matching quantities described in both texts, the author takes this as compelling evidence that Hartlib's manuscript was Newton's source, and calls it a genuinely new finding that may merit publication given Newton's stature, though it has not been submitted anywhere yet. Separately, GPT-6 (unspecified variant) was mid-search through Darwin's notebooks and correspondence, an idea for the search reportedly suggested by the model itself, looking for undiscovered links between Darwin and his informants on natural selection, including a note about a kangaroo larynx; the work was still in progress with no result yet. As the post was being written, Opus 5.5 was also partway through deciphering two 16th century Spanish letters written in secret code, an effort the source text cuts off before describing the outcome.
The author frames all this as a case for AI labs, historians and funding agencies to fund structured collaborations, arguing the approach works best on "tractable" problems: ones where experts have already identified open questions, the data is fully digitized, the task suits models' strengths in multilingual reasoning, math and autonomous search across large corpora, bespoke code can be written to attack it, and any proposed solution can be clearly proven or disproven.
Key facts
- GPT-6 Astra helped break a July 10, 1941 Enigma message by surfacing an overlooked 2026 note about newly found Bundesarchiv radio message collections; researcher Frode Weierud says his team is still analysing how the model located some referenced files.
- GPT-6 Astra concluded John Dee's coded manuscript Liber Loagaeth is mostly nonsense syllables rather than real cipher, but found one passage that does encode a reference to the angelic figure Bornogo.
- Opus 5.5 downloaded over 5,000 primary source files from Samuel Hartlib's archive and found that Newton's anagram 'Vltimorui' and Hartlib's 'Miloirtiua riciragnun' both encode 'vitriolum,' suggesting Hartlib's manuscript was Newton's source.
- Opus 5.5 was mid-way through partially deciphering two 16th century Spanish coded letters as the piece was written.
- The author argues AI labs, historians and funders should pursue collaborations, naming codebreaking, cross-translation tracing, and linking siloed findings as the most tractable problem areas.
Why it matters
The piece argues that frontier models have crossed a threshold: rather than just transcribing or summarizing documents, GPT-6 Astra and Opus 5.5 are being used to attempt actual solutions to open historical problems, from codebreaking to source attribution, something the author says was not realistic as recently as last year.
Who it affects
Historians, archivists and cryptology researchers such as Frode Weierud, along with the teams behind archival projects like the Darwin Correspondence Project and Hartlib's papers; the author's pitch is specifically aimed at AI labs and funding agencies to back paired historian-model collaborations.
How to use it
The author lays out five criteria for picking tractable problems: experts must have already identified the open questions, the source data must be fully digitized, the task must fit models' strengths in multilingual reasoning, math and autonomous search, it must be amenable to bespoke code, and any proposed solution must be provable or disprovable. On that basis he flags codebreaking, tracing texts across translations and adaptations, and connecting findings scattered across niche subfields as the most promising areas.
How solid is it
Several results are explicitly preliminary: the Darwin search is still in progress with no findings yet, the Hartlib report is described as preliminary, the Newton/Hartlib anagram finding is said to 'may' merit publication but has not been submitted, and the piece cuts off mid-sentence while describing the still-unfinished decipherment of two Spanish letters. Even Weierud, the human expert on the Enigma break, says his team is still analysing GPT-6 Astra's logs and does not fully understand how it accessed some of the files it referenced.
Risks and caveats
The author himself cautions that not every result is a breakthrough: the Liber Loagaeth analysis, while interesting, is called 'not a meaningful breakthrough' in Dee studies, and a search for new WW1/WW2 era ciphers to crack came up empty. Weierud also flags that a private collection cited by GPT-6 Astra in the Enigma case has an unclear source, underscoring that these models can push conclusions further than even expert historians can fully trace.
“The file references GPT-6 Astra mentions, RS 3-3/20a and RS 3-3/63b, are correct, but they are not available on the Crypto Cellar Research website. GPT-6 Astra mentions a private collection, but it is not clear what this is, whether it has succeeded in accessing the Bundesarchiv's digitised collections or whether it has found these files elsewhere.”
— Frode Weierud, historical cryptological researcher