Here's a roundup of today's topics, drawn from additional data collected on the evening of September 17. The headline grabber is a report that a Nazi German radio message unsolved for 83 years was deciphered in 10 hours using GPT-6 Astra. At the same time, warnings are mounting against the “just add more AI agents” mindset — and set against OpenRouter's weekly consumption statistic of 126 trillion tokens, it made for a telling day. On the business side, Cohere and Aleph Alpha have merged; in semiconductors, Huawei has spoken out on AI chip supply and demand; and in local LLM land, the first M5 Ultra benchmarks appeared alongside results from running agents on a real smartphone.
GPT-6 Astra Reportedly Decodes Enigma Message Unsolved for 83 Years in 10 Hours
A Bloomberg developer claims to have used OpenAI's GPT-6 Astra to decrypt an Enigma-encrypted radio message sent by the German Wehrmacht in 1941. The 82-character message had gone unread for 83 years, and its contents were reportedly a terse inquiry from a soldier asking about his own marching route. The decoding is said to have taken about 10 hours.
That said, treating this result as settled fact would be premature. The report itself explicitly notes that the decryption has not yet undergone independent verification, so we'll need to wait for the cryptography community to check it. Even so, this kind of task — not brute-forcing the keyspace, but having the model build hypotheses from context and the operational procedures of the era — is precisely the sort of thing where frontier models with increasingly strong long-form reasoning are showing their strengths, making this a symbolic showcase of that progress. Cracking Enigma during World War II took a constellation of brilliant mathematicians and early computers; 83 years later, that cipher is being tackled by a single commercial model. How the verification plays out will be worth watching closely.
More Agents Isn't Automatically Better — The “Coordination Tax” and 126.2 Trillion Tokens a Week
Eric Provencher, a developer on OpenAI Codex, is sounding the alarm on agent sprawl, backed by some eye-popping cost figures. According to him, running more than two parallel subagents almost always just burns tokens without improving quality. The culprit is the “coordination tax”: agents can't trust each other's output, so they end up double-checking one another's results. The cited example is a project where 1,393 agents burned $20,000 worth of tokens on a single Python refactoring job — work that a single Astra agent could presumably have handled at a tiny fraction of the cost.
OpenRouter statistics reported the same day provide the backdrop for this warning. Weekly token consumption through OpenRouter has skyrocketed from 0.5 trillion in January 2025 to 126.2 trillion — a surge of more than 25,000%. The article's analysis is that this figure reflects less a growth in actual usage than the rise of reasoning models that consume enormous token counts per response, multiplied by AI agents being mass-produced without optimization. It's a number that compresses the “AI bubble or not” debate into a single chart — but flip it around, and it also reads as data backing the orthodox conclusion that polishing the design and instructions of a single agent beats scaling up the headcount, on both cost and quality.
Cohere and Aleph Alpha Merge — Taking Aim at the Enterprise AI Market
According to Reuters, Canada's Cohere and Germany's Aleph Alpha have merged, with the stated aim of boosting their presence in the enterprise AI market. Both companies have been outgunned in general-purpose chat AI by giants like OpenAI and Google, yet both are players with solid customer bases built on enterprise data integration and sovereignty and security requirements. The detailed terms and structure of the merger were not part of the data collected this round, and further reporting is expected to follow. The fact that two mid-tier players from Europe and Canada are teaming up hints at a broader realignment: the axis of resistance to the giant model companies is consolidating around enterprise specialization.
Huawei: “AI Chip Demand Exceeds Supply” — Stepping Up the Challenge to Nvidia
Reuters reports that Huawei has stated that demand for its AI chips exceeds supply. The company was also reported to have announced two new AI chips, signaling a further escalation of its challenge to Nvidia under US export controls. This connects to the story from our last digest about a 29B-parameter model trained entirely on Chinese Ascend NPUs: within China, AI chip scarcity is becoming a constraint on capability improvements, while the sheer scale of that demand is itself turning into a business opportunity for Huawei.
First M5 Ultra Benchmarks Make the Rounds — 50 Tokens/sec Generation on a Quantized 27B
A post gaining traction in Reddit's LocalLLaMA community points to Apple M5 Ultra benchmark results published on omlx.ai. According to the results, a 4-bit quantized Qwen 3.8 27B delivers roughly 50 tokens/sec generation at 8k context, with prefill (prompt ingestion) around 1,800 tokens/sec. These are numbers without MTP (multi-token prediction). The poster themselves hedges with “unclear whether these are official, but they look plausible,” so the usual caveats apply. Still, if a 27B-class model runs at this speed, a top-spec Apple Silicon configuration makes running large models on a single desktop an even more compelling option for local LLM setups.
Can an AI Agent Survive a Day on a Real Smartphone? — First Measurements from “AndroidLife”
The first results from a new benchmark called “AndroidLife” have been released. The design runs 60 real-world tasks back-to-back on an everyday OnePlus device (a physical handset connected over Wi-Fi, not an emulator), scored by the device's actual state rather than the model's self-reporting. First up was Alibaba's Qwen3.8-27b (text mode), with a success rate of 56.7%. That's the best text score on the current leaderboard, but it still failed on 43% of the tasks.
Some of the granular numbers are fascinating: an average of 29.25 steps, about 6 minutes, and $0.118 per task. Battery drain was 69%, and the chip temperature peaked at 98.2°C. By difficulty, the results trace a clean decay curve: 80.8% on easy tasks, 52.9% on medium, and 23.5% on hard. The model handles within-app operations reasonably well, but failures reportedly clustered in cases spanning three apps, where it would spin its wheels until hitting the step limit.
Even more telling is the quality of the failures. Of 11 tasks where the model should have asked the user a question, it actually asked only twice. On 4 of 7 “hallucination control” tasks with deliberately missing data, it honestly answered that the data doesn't exist rather than fabricating. Meanwhile, 3 tasks it claimed as successes were rejected on device verification (calling overlapping calendar entries conflict-free, answering a travel time without ever opening the relevant app, and picking a name without confirming). The early data sketches a clear picture: the bottleneck for smartphone agents isn't raw model intelligence so much as navigation across multiple apps and the judgment of when to ask — that's the structural takeaway.
Wrap-up
The Enigma decryption report (GPT-6 Astra), the “coordination tax” warning on agent sprawl alongside OpenRouter's 126 trillion tokens, the Cohere–Aleph Alpha merger, Huawei's AI chip supply-and-demand comments, the first M5 Ultra benchmarks, and agent measurements on a real smartphone (AndroidLife). No flashy new model launches, but a day of news that kept asking the same question: how to design AI usage and its costs. See you at the next digest.