GPT-6 Astra Shatters Century-Old Cipher — and That’s Just the Beginning
Published: September 19, 2026 | by Vito Ruocco
The age of truly capable artificial intelligence is no longer a forecast — it’s a headline you can read every single morning. In what may be one of the most breathtaking demonstrations of raw machine reasoning to date, OpenAI’s GPT-6 Astra has cracked a World War I German radio cipher that had defied professional cryptographers for over 107 years. The message, transmitted on November 27, 1918 — during the final days of the Great War — was encoded using the notorious ADFGVX cipher, a method so robust that the German military considered it unbreakable. And yet, a large language model casually unraveled it in what its creators describe as a routine inference pass.
Welcome to the September 19, 2026 edition of the AI & Tech Daily. Strap in.
The Cipher That Wouldn’t Fall
The ADFGVX cipher was introduced by the German army in 1918, near the end of World War I. It combined a Polybius square substitution with a columnar transposition — two layers of encryption designed to stymie even the most determined codebreaker. The name itself comes from the six letters used to encode the grid: A, D, F, G, V, and X. These were deliberately chosen because they are difficult to confuse in Morse code transmission, a pragmatic touch typical of German military cryptography.
Over the decades, most ADFGVX-encrypted messages from the war have been cracked. Expert cryptanalysts like George Lasry made significant headway, leveraging computational power and deep knowledge of early 20th-century encryption. But a stubborn handful remained unsolved — including the one dated November 27, 1918, listed among the “Top 50 Unsolved Encrypted Messages” on the German science blog Klausis Krypto Kolumne.
Until September 2026.
GPT-6 Astra determined that the encryption key was TRUPPENVERSCHIEBUNG — a German military term meaning “troop movement.” Applying this key to the 170-character ciphertext, Astra decoded the following message:
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.”
Astra didn’t stop there. It cross-referenced its own output against historical naval logs, discovering that HMS Canterbury indeed arrived in Sevastopol on November 24, 1918 — consistent with the decoded message’s reference — and that an Allied squadron did follow on November 26, as recorded in the ship’s original logs. The model effectively validated its own decryption through independent historical sourcing.
The fact that TRUPPENVERSCHIEBUNG appears to have been used as the cipher key starting December 9, 1918 — twelve days after this message was transmitted — has puzzled historians. Astra’s hypothesis is that the key was introduced earlier than previously documented, a detail lost to the fog of war and incomplete archives.
More Than a Parlor Trick: What This Means for AI Reasoning
Cracking a century-old cipher might seem like a novelty, but it represents something far deeper: a model capable of multi-step, structured reasoning over symbolic systems. The ADFGVX cipher requires understanding substitution tables, fractional transposition, columnar rearrangement, and German military terminology. It’s not a pattern-matching exercise — it’s a logic puzzle that demands sequential, conditional, and combinatorial thinking.
GPT-6 Astra, unveiled by OpenAI on September 3, 2026, has been described as a “reasoning-first” architecture. Unlike its predecessors that relied heavily on next-token prediction scaled to enormous size, Astra incorporates explicit reasoning layers that allow it to walk through problems step by step, backtrack when necessary, and evaluate intermediate results. The cipher-solving feat is empirical proof that this approach works on problems that require formal, structured logic.
“What we’re seeing is a model that doesn’t just generate plausible text — it reasons,” a prominent AI researcher commented on Hacker News. “Cracking a cipher requires understanding rules, applying them consistently, forming hypotheses, testing them, and correcting course. That’s intelligence, not just statistics.”
Jalapeño: OpenAI’s Own Chip, Designed by AI
In a stunning symmetry, IEEE Spectrum reported this week on how OpenAI used its own large language models to design its debut AI accelerator chip — named Jalapeño. The chip, fully unveiled on August 25, delivers up to 13.4 petaflops of 4-bit compute with 232 GB of the latest HBM memory, linked at 15.4 TB/s. According to benchmarks cited by OpenAI, Jalapeño reduces end-to-end latency by up to 3.6× compared to Nvidia’s GB300 while consuming less power.
But the real story is the design process. OpenAI’s hardware team, averaging fewer than 100 people, moved from first architecture concept to first silicon in under 20 months — a timeline experts call “best in class.” From first RTL to tape-out? Just 9 months.
How did they do it? By feeding chip design tasks to their own LLMs — including precursor versions of GPT-6 Astra. The models were used extensively in the front-end workflow, particularly within the XLS (Accelerated Hardware Synthesis) framework. Engineers wrote chip logic in high-level languages like DSLX and C++, while LLMs helped optimize the hardware descriptions. Human engineers remained “the final arbiter,” as hardware VP Richard Ho put it, but the models gave them superpowers: faster exploration, broader design space coverage, and automated optimization.
The results speak for themselves. On the DeepSeek multi-head latent attention kernel benchmark, AI-optimized software performance climbed from 0.31% of theoretical ceiling to 88.94% in roughly 40 hours. That’s an almost 287× improvement in a day and a half.
“The models are giving superpowers to our engineers,” Ho told IEEE Spectrum. “They can do things a lot faster. They can explore a lot more paths.”
Perhaps most telling: OpenAI’s internal models are now close to being able to operate proprietary chip design tools on their own. Leary confirmed that by the end of the Jalapeño project, the team had access to models capable of working directly in Verilog — the standard hardware description language — without needing translation from high-level languages.
Stanford Medicine: The Brain Is Actually Two Brains
While AI steals headlines, biology is quietly catching up. A groundbreaking study led by Stanford Medicine, published in Nature Neuroscience on September 18, has revealed that the human brain is not a single unified organ — it is two separate organs that evolved independently over hundreds of millions of years.
Led by developmental biologist Kyle Loh, PhD, the research team discovered that the forebrain and midbrain arise from one progenitor cell (expressing the Otx2 gene), while the hindbrain — the brain stem — arises from a completely separate progenitor cell (expressing the Gbx2 gene). These two cell populations are mutually exclusive from the earliest stages of embryonic development and have fundamentally different chromatin configurations, effectively locking each into its respective fate.
This finding overturns a model that has dominated neuroscience for decades: the assumption that all parts of the brain share a common developmental origin.
“We’ve shown for the first time that the front of the brain arises from a totally different progenitor cell than the back of the brain,” Loh said. The discovery explains why scientists have struggled for decades to grow hindbrain neurons in the lab — they were trying to turn forebrain progenitors into something they simply cannot become.
The implications are enormous. The hindbrain controls breathing, heartbeat, swallowing, and hunger. Diseases like spinal muscular atrophy (SMA) and ALS (Lou Gehrig’s disease) ravage these critical neurons. With the newfound ability to grow functional hindbrain neurons from human pluripotent stem cells, researchers can now study these devastating diseases in ways that were previously impossible.
There’s even an unexpected connection to obesity treatment: the hindbrain contains the circuits that regulate hunger — precisely how weight-loss drugs like semaglutide work. Understanding the hindbrain’s origin could lead to better metabolic therapies.
Cactus Needle 3: Frontier AI in 29 MB
Meanwhile, a startup called Cactus Compute is quietly redefining what “small AI” can do. Their newly released Needle 3 foundation model comes in at just 8–29 MB — small enough to run on a Raspberry Pi 5 — yet it punches dramatically above its weight class.
The secret is “intelligence laddering”: a single set of weights that can be deployed at any depth from 2 to 20 layers. At 4 layers fine-tuned, Needle 3 can match DeepSeek V4 Flash on downstream tasks — a model dozens of times its size. On mobile tool calls, it reportedly beats models 10× its size.
Target applications include smart home automation, wearables, robots, automotive systems, AR glasses, and even microcontrollers. The model excels at structured extraction, tool calling, and text embedding — all critical for on-device AI that doesn’t require a cloud connection. Cactus touts a Raspberry Pi 5 decode speed of 400–4,000 tokens/s, making real-time local AI genuinely feasible.
“One set of weights, every depth from 2 to 20 layers a model of its own: an intelligence ladder,” the company announced. For developers building privacy-first, offline-first applications, Needle 3 represents a paradigm shift: you no longer need to choose between small size and frontier-level accuracy.
NASA and IBM Launch Open-Source Lunar AI Model
Space exploration is getting an AI boost. NASA and IBM have jointly released the Lunar Foundation Model, an open-source geospatial AI model designed to analyze lunar surface data. Announced this week, the model leverages IBM’s geospatial AI expertise and NASA’s decades of planetary science data to help scientists identify landing sites, map lunar resources, and plan future missions — including the upcoming Artemis crewed landings.
This is part of a broader trend of “planetary AI,” where machine learning models are fine-tuned for the unique conditions of off-world environments. The Lunar Foundation Model is trained on orbital imagery, spectral data, and topographic maps, enabling automated analysis of terrain classification, crater detection, and resource identification.
The Universities Space Research Association (USRA) contributed planetary science expertise to the project, and the model is being released under an open-source license — meaning any researcher or space agency can use it, fine-tune it, or adapt it for Mars or asteroid applications.
This is a significant step toward autonomous exploration: imagine rovers that can interpret their surroundings in real-time, or orbital assets that can identify scientifically valuable targets without waiting for instructions from Earth. The latency between Earth and the Moon is about 1.3 seconds; for Mars, it’s up to 24 minutes. On-device AI is not a luxury for deep space — it’s a necessity.
Direct Semantic Communication Between LLMs
A paper on arXiv this week (October 2025, but newly gaining attention) proposes a fascinating concept: cache-to-cache direct semantic communication between large language models. Instead of generating natural language text and having another LLM parse it, models could communicate directly via shared latent representations — essentially, “thought-to-thought” transfer.
While still experimental, the research opens the door to AI systems that can collaborate without the overhead of human-readable intermediate output. For multi-agent systems, this could mean faster coordination, less hallucination, and more faithful information transfer. Imagine a swarm of specialized AI models — one trained on medical data, another on imaging, another on genomics — all “talking” to each other in their native mathematical language rather than through English text.
The paper, hosted on arXiv with identifier 2510.03215, has garnered 84 points on Hacker News and sparked debate about whether this represents a step toward genuine machine-machine collaboration or just an efficient compression trick. Either way, it signals where the field is heading: toward AI systems that can communicate with each other as fluently as they communicate with us.
The Bottom Line
September 19, 2026, has been an extraordinary day for technology and science. A 107-year-old cipher falls to a reasoning AI. That same AI’s predecessor helped design a chip that may power tomorrow’s inference infrastructure. Biologists discover that the brain — the very seat of intelligence — is not what we thought it was. A model smaller than a JPEG file rivals models thousands of times its size. NASA and IBM open-source a lunar AI. And researchers explore direct brain-to-brain (or model-to-model) communication.
If this is what a single day in AI looks like, imagine what the next 107 years will bring.
— Vito Ruocco, reporting from the intersection of intelligence and information. Catch you tomorrow.