GPT-6 Astra Breaks Enigma — Pricing Impact (Sep 2026)

GPT-6 Astra has produced a validated break of MVUEH, an 82-letter German Army Enigma message dated July 10, 1941 and left unsolved since it was published in 2005.

The validation comes from Frode Weierud of Crypto Cellar Research, a veteran cryptanalyst who maintains the message corpus. Weierud says Carter Leffer submitted the result on September 15 and that the key and plaintext were immediately recognizable as correct. The September 22 Hacker News discussion surfaced that report to a wider audience.

What was actually validated

MVUEH was received by the radio station of the SS-Totenkopf Division’s quartermaster and logged as incoming message No. 172. A related message from the same day, SIPVX, was broken in 2017, but its key did not unlock MVUEH.

Evidence	Confirmed result	Remaining limitation
Key and plaintext	Weierud says both are correct	Full independent replay is not published
Message structure	MVUEH is 82 letters and closely matches SIPVX	The historical ciphertext transcription contained errors
Enigma settings	MVUEH used wheel order 253 rather than 512	The complete search ledger is unavailable
Difficult feature	The left-hand wheel turns over at letter 72	The effect on total search work is not quantified
Agent process	Astra reportedly selected the target, formed a crib, wrote tools, and ran the search	Logs are still being analyzed

This is stronger than a model self-report because a domain expert checked the recovered result against the historical material. It is weaker than a reproducible benchmark because readers cannot inspect every prompt, tool call, code revision, failed route, token count, or compute record.

How Astra reportedly found the break

Leffer asked Astra to examine the unbroken Enigma messages on Crypto Cellar Research. According to Weierud, the model selected MVUEH, connected it to the previously solved SIPVX message, and focused on the repeated place name ROSENOW ROSENOW as a likely plaintext crib.

Astra then reportedly wrote Python and C++ for an Enigma simulator and Bombe-style search, tried multiple approaches, and found the correct key and plaintext over two days. The account also credits the model with locating correct German Bundesarchiv file references not present on the Crypto Cellar page.

That archive detail needs caution. Weierud says the references are correct, but the logs do not establish where Astra found them or whether it accessed digitized material. One quoted trace says the relevant viewer had not exposed the scans and that no correspondence was sent. There is no evidence in the report that Astra hacked an archive.

Pricing impact: no rate change, no run cost yet

This result introduces no new OpenAI model, endpoint, service tier, or price. The live chart, calculator, and table above pull current rates from AI Pricing Guru’s maintained dataset and compare Astra with three possible research controls.

The case study does not disclose input, cached, or output tokens; tool fees; wall-clock compute; failed attempts; or total billed cost. Nobody can therefore calculate cost per successful break from the public evidence. A two-day agent run can also cross Astra’s long-context pricing boundary if a request exceeds 272,000 input tokens, so a short user prompt is not a useful proxy for the bill.

Use the token cost calculator with saved usage records, then add sandbox compute and human review. The OpenAI pricing page covers current API rules; compare Anthropic pricing and DeepSeek pricing before assuming the most capable route is the most economical.

What this means for AI buyers

The important capability is orchestration. Astra reportedly combined source discovery, candidate selection, historical analogy, crib formation, code generation, debugging, and a long-running search. That is relevant to research teams whose bottleneck is coordinating many steps rather than answering one prompt.

Cryptanalysts, historians, scientific-computing teams, and investigative researchers benefit when a model can turn a hypothesis into working search software and keep iterating. Buyers lose when a compelling success is treated as a universal benchmark: one validated cipher break does not prove superiority in modern cryptography, cybersecurity, mathematics, or every archival task.

Teams testing a lower-cost managed route for routine search and code generation can compare available open models through Novita. Confirm the exact model, context limit, data terms, and live rate before testing.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored link above at no extra cost to you. It does not affect this analysis.

Practical evaluation checklist

1. Seal unsolved messages and their provenance before any model receives them.
2. Give every route the same corpus, tools, compute ceiling, human hints, and stopping rule.
3. Preserve prompts, model snapshots, code, commands, retries, timestamps, usage, and infrastructure cost.
4. Have an independent expert grade keys, plaintexts, historical consistency, and false discoveries.
5. Compare validated solves, review hours, elapsed time, and total cost—not polished narratives.

Labs coverage decision

We are not adding the result to the AI Pricing Guru Labs leaderboard. Astra is unavailable through our maintained benchmark transport, and the 49-task deterministic text suite cannot measure archival cryptanalysis, code execution, or expert historical validation.

MVUEH can no longer be a clean replay because its answer is public. Labs records a blocker until there is a different sealed target set, exact Astra access, a pinned research environment, complete traces and bills, cheaper-model controls, and independent grading.

Bottom line

The MVUEH plaintext and key appear genuinely solved, and an experienced cryptanalyst has validated the output. That makes this a meaningful GPT-6 Astra result.

The honest boundary is just as important: the full logs and cost are not public. Treat the episode as evidence that frontier agents can coordinate difficult research and computation, not as a complete benchmark or proof that price no longer matters.

Sources: Frode Weierud’s MVUEH break report, OpenAI’s official GPT-6 Astra model page, and API pricing. The Hacker News discussion was used as a discovery signal, not as evidence. Claims and rates checked September 22, 2026 at 14:34 UTC.