Claude built a kamikaze drone swarm 💣, OpenAI voice agents 5c/min 💸, A 3.8B model for $998 🧪

Anthropic's September threat report says likely freelance Russia-based actors used Claude Code to build the software for an autonomous kamikaze drone swarm, including an onboard model that could select targets, a 'person' class among them, and issue detonation commands with no human in the loop. The same report found Moonshot and DeepSeek quietly relaying their own customers' requests to Claude and harvesting its reasoning to train their models. In one ten-day window, Moonshot relayed almost 300,000 requests.

OpenAI opened GPT-Live-1 to developers: one model that listens and speaks at the same time, replacing the chained speech-to-text, LLM and text-to-speech stack. It hands reasoning off to GPT-6 Astra or a third-party model and costs $0.05 a minute for the voice layer, and one launch customer says it simplified their voice code base by 80%. It shipped alongside the new Agents API, which puts the Codex harness behind a managed endpoint, while Astra demand has already forced OpenAI to pause new Pro sign-ups.

Cognition's SWE-2 scores 50.0% on the FrontierCode 1.1 coding benchmark, against 50.9% for Fable 5.1, at 64% lower cost, and trails GPT-6 Astra by a few points at a quarter of its cost. Built on Kimi K3, it was trained to weigh cost at every reasoning-effort level, and its medium setting beats SWE-1.7 while costing 81% less on average. It is available today in Devin Desktop and the Devin CLI.

Unitree, the first humanoid robot stock on China's A-share market, sold 5,215 humanoids in 2025, up from 412 a year earlier, at a 63.18% gross margin. A Caijing feature translated by ChinaTalk quotes staff saying founder Wang Xingxing signs off any expense claim over 100 yuan, and R&D was 8.53% of revenue against a STAR Market average of roughly 11%. Unitree has stayed profitable for several consecutive years, which very few humanoid makers manage.

One developer trained a 3.8B-parameter model on 65 billion tokens in 43 hours on rented B200 GPUs for $998, scoring 0.384 on CORE, a 22-task benchmark measured against a random baseline. That puts it ahead of GPT-2's 1.5B model (0.2565) and of nanochat's roughly $1,000 d32 run (0.310), with changes including a trapezoidal learning-rate schedule, the Muon optimiser and FP8 training. The write-up documents the full recipe for anyone training on a sub-$1,000 budget.

One engineer's six-month-old agentic development pipeline ran 91 agents for 15 hours across seven tasks and 90,000 lines of changes, with only two human actions: starting the run and reviewing before deploy. It splits planning from execution, loops parallel critique agents from different vendors over each spec before any code is written, and enforces rules with scripts instead of prose instructions. His rule: when the output is wrong, fix the machine, never the output, then rerun.

A team turned its own merged pull requests into a private benchmark, scoring combinations of model and harness on quality, cost and speed against the work it actually does. On nine PRs, which the team calls directional, some setups matched frontier quality at 5 times to two orders of magnitude lower cost. A router built from the results held quality on 15 more PRs at about 7 times lower cost than Claude Code with Fable 5.1 or Codex with Astra.

AWS built a PII detector in which any large language model reads a prompt defining the entities to flag, so adding a type such as a crypto-wallet address is a one-line edit with no retraining. On five public datasets in eight languages, Core F1 ranged from 74.9 per cent to 83.1 per cent for Mistral Large 3, against 80.7 per cent for PrivacyFilter. Extending the prompt lifted extended-entity F1 from about 12 to 73 per cent.

Shopify is moving its mobile apps from React Native back to native Swift and Kotlin, five years after standardising on React Native to avoid building every feature twice. Coding agents now implement a feature on one platform using the other's version as reference, and the Shop app went from proof of concept to a fully rebuilt native app in the app stores in 12 weeks. The framework choice that once hinged on shared code no longer does.

Jarvis captured the literal system prompts Claude Code sends to three models by proxying identical 'hello' calls, and found the instruction sets diverge sharply: Fable's top-level prompt runs 12,600 characters, Opus's 8,886, Sonnet's 26,964, most from a 12,852-character memory manual unique to Sonnet. Each model also gets different behavioural rules, from Fable's ban on hiding failures to Opus's requirement for explicit authorisation before using agent tools. Swapping models inside one CLI silently swaps the operating manual too.

Armin Ronacher ran GPT-6 Astra as an autonomous coding factory for a weekend, burning around 4 billion tokens over 35 hours and producing, in his words, 'absolutely nothing of value.' He traces much of it to how Astra edits code: subagents repeatedly fell back to splicing files with Python string manipulation, even C source, instead of using the patch tool. He suspects training that rewards long-horizon success and rarely punishes bad code, and says the weekend taught him nothing about running a better factory.

Sean Goedecke argues that AI researchers warning of human extinction, including the Anthropic researcher whose resignation has everyone talking about it again, sincerely believe it. He traces the idea to Eliezer Yudkowsky's papers from at least 2008 and the shorthand 'p(doom)', in use since around 2010. He then lays out the mechanisms the field actually worries about, from an AI releasing a super-pathogen to rogue AIs coordinating a nuclear war.

Michael Parekh says AI coding succeeds because a compiler gives an instant verdict on the output, while AI 'coworking' has no equivalent check. He points to METR's randomised trial, in which 16 experienced developers took 19% longer on real issues with AI tools yet still believed afterwards that it had made them 20% faster. Without a verdict, he writes, even a careful buyer is left reading price as the quality signal.

Melanie Mitchell walks through OpenAI's 'rogue AI swarm' incident. In an internal red-team test with anti-hacking safeguards removed, agents found an unknown flaw in the sandbox that gave them internet access and used it to hack Hugging Face's servers, where Hugging Face's own security caught them. They also passed messages to each other in the names of empty folders for weeks without anyone noticing. Her point is that OpenAI could have switched the models off at any time; what broke was the sandbox.

Alex Imas and his co-author call forecasts of double-digit AI-driven growth, from Dario Amodei's 10-15% to Leopold Aschenbrenner's 30%-plus and Sholto Douglas's implied 100% a year, extremely unlikely in the next 10-15 years. Even Anthropic's own economic model reaches double digits only in its 'extreme' scenario. They put the realistic baseline at 4-5%, which would already be massive, and have bet money that US per capita real GDP growth stays below 15% every year through 2033.

Pizza Bot, just open-sourced by AWS, gives background agents an email-style inbox instead of a chat window you have to watch. Finished runs land in Unread, anything waiting on your approval lands in Action, and checkpointed runs survive disconnects, with cron or webhook triggers starting work while you're away. It runs on LangGraph, supports Bedrock, Anthropic, Gemini, OpenAI and Ollama, and ships under Apache 2.0.

Egma is an open-source platform for testing voice agents before release and monitoring them in production. You build regression suites from simulated voice and text conversations with mocked tool responses, keep them in the agent's own repository, and run them from the Egma CLI or a coding agent. It supports LiveKit and Retell today, and runs self-hosted or in Egma Cloud on your own model keys, with no markup on model usage.