🎬 Meta Muse ships 10-second AI videos with native audio 🎬
Meta's closed-beta video model rolls out with built-in audio generation.

Meta's Muse Video model has entered closed beta, producing 10-second AI-generated videos with native audio — no separate sound model required. Early samples show strong fine detail, solid world understanding, and impressive temporal consistency across frames. The inclusion of integrated audio is what sets Muse apart from most competitors in the space, which still rely on separate pipelines for sound. This matters because it could reshape content creation across marketing, entertainment, and social media. The model is not yet publicly available, but closed-beta outputs already circulating online suggest the quality is competitive. If it holds at scale, Muse could compress creative video workflows that currently take hours into minutes.

🔒 OpenAI adds private safety to zero-retention API 🔒
New system scans API sessions for misuse while keeping data invisible to humans.

OpenAI introduced Private Safety Processing, a new architecture that scans API sessions for abuse patterns without any OpenAI employee ever seeing the underlying content. The system preserves the company's zero-data-retention commitment while still enforcing automated safeguards — a direct response to the tension between safety and privacy that has kept many regulated organizations from adopting frontier models. Available now in preview, the feature targets finance, healthcare, and government clients on zero-retention plans. The competitive pressure this creates is real: enterprises now have fewer reasons to avoid OpenAI's API on compliance grounds, potentially unlocking adoption in sectors that previously defaulted to on-premise solutions.

🧬 Claude designs proteins at 35% hit rate 🧬
Anthropic's models autonomously ran protein-design campaigns, replacing days of lab work.

Anthropic published research showing that Claude Mythos Preview and Opus 4.8 autonomously ran protein-design campaigns with minimal human guidance, producing working molecules on 14 out of 15 targets. The models achieved 22–35% hit rates for target-binding molecules, far exceeding the industry-standard 10–15%. Separately, Opus 5 analyzed raw lab instrument files without specialized software, measuring sample purity at 96.4% in just 19 minutes — compared to a four-day turnaround using standard lab processes. Lab partners included Twist Bioscience and Adaptyv Bio. This signals a capability jump for AI-driven drug discovery and positions Anthropic as a serious contender in life sciences, not just coding and reasoning.

INTERESTING TO KNOW

🖥️ Meta AI desktop app lands on macOS with screen sharing 🖥️

Meta released a standalone macOS desktop application for its AI assistant, rolling out real-time screen-sharing capabilities that let the AI see and understand what users are working on. This puts it in direct competition with Anthropic's Claude and OpenAI's ChatGPT desktop clients for professional integration. Screen sharing enables contextual assistance — referencing on-screen content for coding, writing, or analysis — signaling Meta's intent to embed its AI deeply into desktop productivity workflows beyond its social and consumer roots.

⚡ Replit ships Free Mode via GPT-5.6 Luna ⚡

Replit launched Free Mode, a pricing integration that routes everyday chats and routine tasks to OpenAI's GPT-5.6 Luna at zero credit cost — scaling usage up to 30x on the $20/month Core plan with an additional 30 hours of monthly chat included. Complex builds still consume credits, and the agent can auto-escalate to OpenAI's Sol model for harder steps. This tiered approach commoditizes basic AI interactions while monetizing advanced ones, substantially lowering the access barrier for solo developers and small teams to use AI-assisted coding around the clock.