🤖 Anthropic ships Fable 5.1 with record benchmark score 🤖
Anthropic's rollout doubles prior benchmarks and opens tiered access for frontier researchers.

Anthropic launched Claude Fable 5.1, the successor to Fable 5, which now sits atop Artificial Analysis' Intelligence Index with a record score of 66 — more than doubling its predecessor on scientific research benchmarks. The model delivers major improvements on long coding tasks and research workloads. Critically, safety filters now intervene 60% less on cybersecurity work and 85% less on basic medical and biology queries, directly addressing one of the biggest customer complaints. Alongside it, Anthropic released Mythos 5.1, the same underlying model but with fewer guardrails, available exclusively to screened U.S. cybersecurity and biology researchers. This tiered-access approach reshapes how frontier capability reaches different user groups, signaling a competitive pressure shift across the industry.

🎙️ Meta drops Muse Voice Transcribe for live audio 🎙️
Meta's new model handles real-time speed across 20+ speakers at once.

Meta released Muse Voice Transcribe, its first real-time audio perception model capable of converting live audio streams into text while distinguishing more than 20 speakers simultaneously. The model supports streaming speech recognition, speaker diarization, endpointing, multilingual code-switching, and contextual biasing — and already tops the Artificial Analysis leaderboard for audio-to-text tasks. This directly expands Meta's distribution in real-time multimodal tooling, with applications spanning meetings, live events, and accessibility. Muse Voice Transcribe is available through Meta's research AI channels and third-party tool directories, positioning Meta as a serious competitor in an arena previously dominated by specialized transcription providers.

🛡️ OpenAI's Astra hits 'Critical' cyber rating 🛡️
OpenAI's Astra is the first model to cross its internal regulation-triggering cyber threshold.

OpenAI announced that its upcoming model, codenamed Astra, is the first to cross the company's internal "Critical" cybersecurity capability threshold. The model can reportedly discover previously unknown security vulnerabilities and exploit them autonomously without step-by-step human guidance — a capability jump that raises fresh questions about responsible deployment. OpenAI has restarted the training run it had frozen following the Hugging Face breach, and a limited release is expected soon. Early outputs have already surfaced via testing channels, drawing serious attention from the security community. This marks a clear escalation in AI-driven offensive cyber capabilities.

INTERESTING TO KNOW

💻 Meta launches Muse Code to rival Claude Code 💻

Meta launched Muse Code, a new coding agent for terminal and CI environments that can plan, edit files, and execute commands — rolling out with security-conscious defaults including mandatory approvals and an OS-level sandbox enabled out of the box. Alongside the release, Meta made Muse Spark available on the Meta Model API and integrated it into the Muse Code workflow. The launch puts competitive pressure directly on Anthropic's Claude Code and similar agentic developer tools, expanding Meta's developer ecosystem well beyond open-weight models.

🪥 Dyson rolls out £499 CameraJet AI toothbrush 🪥

Dyson has rolled out CameraJet, a £499 electric toothbrush with an integrated AI-powered camera that identifies gaps between teeth in real time, then triggers a targeted jet spray for better coverage. The pricing positions it at the top end of the oral care market, but the product represents one of the most consumer-facing deployments of embedded AI vision in a personal care device, signaling growing adoption of AI hardware in everyday consumer health.