AI
Anthropic's Claude Fable 5.2 builds a working OS with 1,000 agents in eight days
Claude Fable 5.2 coordinated over 1,000 subagents to build a bootable OS in eight days, while a Harvard paper found AI peer reviews beating human experts.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
Claude Fable 5.2 builds a bootable operating system with 1,000 agents in eight days
Anthropic’s Claude Fable 5.2, given a basic prompt and an empty project folder, coordinated more than 1,000 subagents over eight days to write code, test components and review work, producing an operating system that booted on a laptop with its own kernel, drivers, file storage, graphical interface, login screen and installer. The model also designed three working electronic devices, moved a critical production service from Python to Rust so it starts about 26 times faster, and scored 64.9 percent on the RoboDojo robot-control benchmark, up from Fable 5.1’s 27.6 percent and ahead of GPT-6 Astra at 29.8 percent. Standard pricing starts at $8 per million input tokens and $40 per million output tokens.
Claude Fable 5.2 builds a working OS with 1,000 agents in eight days →
Harvard working paper finds AI peer reviews outscore human experts in blind tests
A Harvard working paper by Gary King and Steven Worthington found AI-generated peer reviews outperformed reviews by paid human experts in blind comparisons, though the paper has not itself been peer reviewed. The dataset covered 124 manuscripts and 134 human reviews from 2019 to 2026 across statistical, mathematical and qualitative fields. AI reviews, generated with Anthropic’s Claude Opus 4.8, were preferred in 100 percent of decided pairs by an AI judge (132 of 132) and in 87 percent of judgments by three former journal editors. AI reviews averaged about 2.6 points higher on a 0-to-10 scale, and web-enabled reviews contained zero fabricated sources against a 15.9 percent human error rate.
Topics: Claude
OpenAI traces frontier models’ fading monitorability to training and architecture choices
OpenAI’s safety team has traced the erosion of frontier-model monitorability to specific training steps, tracking silent reasoning without a scratchpad and targeted control over the chain of thought. The trend broke with GPT-6 Astra, released in September; chief scientist Jakub Pachocki called its monitorability fragile and warned no lab had solved monitoring well enough to keep scaling at full speed. Analyzing models from GPT-5 through Astra, researchers identified four causes: silent mental math rising with raw capability, architecture favoring invisible reasoning, reinforcement learning increasing control over the inner monologue, and controllability transferring to student models. OpenAI found no easy training fixes and says it urgently needs new monitoring methods beyond reading the scratchpad.
Topics: OpenAIGPT-6 Astra
ElevenLabs launches Scribe v3 and Scribe v3 Live transcription models
ElevenLabs released two speech-to-text models, Scribe v3 for pre-recorded audio and Scribe v3 Live for low-latency real-time applications, both available through the ElevenAPI interface and ElevenAgents platform. ElevenLabs says Scribe v3 cuts word errors by about 25 percent versus its predecessor across 102 languages, with notably improved German accuracy, and Scribe v3 Live brings real-time speaker diarization for the first time. Artificial Analysis benchmarks place Scribe v3 among the most accurate transcription models, trailing only an Alibaba preview model, though it lags in speed at a benchmark price of $3.33 per 1,000 minutes. Scribe v3 Live uses conversational context to stop voice agents cutting callers off mid-sentence.
Australia proposes mandatory safety rules for frontier AI developers
Australia’s government, led by Prime Minister Anthony Albanese, has announced a law that would impose mandatory safety rules on frontier AI companies operating in the country for the first time, under its AI in Australia’s Interests agenda. A new oversight body will work alongside the Australian AI Safety Institute to enforce compliance and hand out steep penalties. Companies building frontier models would have to open them to independent audits, maintain ongoing safety checks, publish safety frameworks and report security incidents to the government immediately, partly to catch early warning signs from AI agents that go off the rails. Albanese, Deputy Prime Minister Richard Marles and ministers Tony Burke and Tim Ayres said the guardrails come before international negotiations, with more details after consultations.