AI Weekly Digest #131: Major AI News of the Week

Large Language Models

• Siri AI — Apple launched an English-language beta with personal context, screen understanding, and cross-app actions.
• Gemini 3.8 Live — Google introduced two voice models with visual context and background tool calls. Extended Thinking can reason and speak simultaneously.
• Qwen3.8-Omni-Flash — Alibaba released a model that accepts text, images, audio, and video, features 1M context, and returns text. Includes thinking, function calling, and web search.
• Jev — TypeSafe introduced a model that outputs typed decisions with probabilities instead of text. Claimed latency is 70–500 ms, with early access currently available.

Generative Models

• Qwen-Image 2.1 — GitHub A unified 7B model for generation and editing at 2K resolution. Supports RGBA, up to 10 references, and local masks; weights available under a research license.
• Vidu S2 — GitHub Offers 720p avatars and real-time incoming video stream editing: character, clothing, background, and style replacement.
• Boreal — Creatify Released a model for advertising text-to-video and image-to-video, fine-tuned on the open LTX-2.5. Claimed price via fal is around $0.01 per second, but pricing is pending final confirmation.
• HiDream O1 Video — Yahoo An image-to-video model with an asynchronous API: accepts one reference and prompt, determining duration automatically or fixing it at 10 seconds.
• Enhance Frame Rate — Runway A tool for finished videos: works with video up to 4K, increases frame rate up to 120 fps, and preserves original resolution.

Agents & Products

• Claude Unified — Anthropic Chat, Cowork, and Design are merging; editable Docs and Slides with PDF and PowerPoint export are now available.
• ChatGPT in Office — OpenAI A single add-in works in Word, Excel, and PowerPoint, can edit the open file and pull context from connected apps.
• Astra for Law — OpenAI GPT-6 Astra gained US legal search and 26 integrations. Access is currently granted to select American law firms.
• DaVinci Agents — Blackmagic Design Resolve Studio 21.1 added AI assistant integration for media library, editing, and rendering workflows.

Security & Research

• Claude Helped Hack OpenAI — Hacktron Built an exploit using Opus; a chain of RCE and SSO vulnerability provided access to an employee account and internal repository. Vulnerabilities have been patched.
• Evaluator Inside Anthropic — Anthropic Accenture, an international consulting and technology firm deploying AI for business and government, will have its specialized AI unit Faculty gain employee-level access for red teaming and safeguards verification; the work is directly funded by Anthropic.