AI NewsSeptember 18, 2026

xAI rolled out voice conversations for Grok Bot, letting users speak naturally to their AI teammates instead of typing, right as its three-day Galaxy livestream event wrapped up.
xAI rolled out voice mode for Grok Bot on September 17, letting users speak naturally to their AI teammates instead of typing. It's powered by Grok Voice Think Fast 2.0, built to minimize the latency that usually makes voice AI feel clunky.
The launch capped xAI's three-day Grok Bot Galaxy livestream, where team members built and demoed business workflows live, emphasizing hands-free talks during commutes or multitasking. Existing users got a usage reset; new livestream sign-ups earned a free month worth $200.
Key Capabilities:

• Natural conversation: Speak to your Grok Bot instead of typing, powered by Grok Voice Think Fast 2.0.
• Low-latency design: Built specifically to cut the delay that usually makes voice AI feel clunky.
• Launch perks: Existing users got a usage reset; new livestream sign-ups earned a $200 free month.

OpenAI launched Astra for Law, a version of GPT-6 Astra paired with a legal search index of over 230 million URLs, scoring 54% on legal research benchmarks versus 39% for the base model.
OpenAI introduced Astra for Law on September 17, pairing GPT-6 Astra with a legal search index covering over 230 million URLs. On Vals AI's Legal Research Bench, it scores 54% correctness versus 38.7% for GPT-6 Astra with web search alone.
Harvey and Sullivan & Cromwell helped build custom workflows on it, and access rolls out to select firms through Trusted Access in ChatGPT and Codex first. API access for Harvey and Legora comes soon, as OpenAI frames this as the start of a longer legal push.
Key Capabilities:

• Legal-grade search: Indexes over 230 million URLs of U.S. case law and statutes, drawing heavily from CourtListener.
• Measurable gains: 54% correctness on legal research questions, up from 38.7% for the base model.
• Firm partnerships: Sullivan & Cromwell, Ropes & Gray, and Cooley built custom tools with OpenAI's help.

Meta released a dedicated Mac app for its personal AI agent Muse, letting it organize downloads, find lost files, and summarize notes across your computer's apps with approval required for sensitive actions.
Meta released a dedicated Mac app for Muse on September 17, letting the agent organize your downloads folder, find a lost file, and summarize messages and notes across your computer. It runs in a secure virtual machine and always asks approval before sensitive actions.
It's currently US-only, with free and paid tiers, building on Muse's iOS, Android, and web versions from earlier this month. Early users give it five-star reviews, calling it the best personal assistant they've tried for everyday tasks like booking or chores.
Key Capabilities:

• Cross-app control: Organizes downloads, finds lost files, and summarizes notes and messages.
• Approval-gated: Always asks before taking any sensitive action on your files.
• More coming: Currently US-only, with additional Mac features planned soon.

OpenArt's new Arena leaderboard has BytePlus's Seedance 2.5 and Seedream 5.0 Pro topping the overall video and image rankings, based on blind side-by-side comparisons from creative professionals.
OpenArt launched Arena this week, ranking AI image and video models by task instead of one overall score. Seedance 2.5 tops overall video and leads film, motion design, and lip sync; Seedream 5.0 Pro wins overall image plus film and e-commerce.
It's not a clean sweep: GPT Image 2 leads graphic design and image editing, and Wan 3.0 edges out Seedance 2.5 on video editing by a single point. Professionals judged two outputs at a time without knowing which model made either.
Key Capabilities:

• Task-specific rankings: Separate leaderboards for graphic design, film, e-commerce, lip sync, and more.
• Split leadership: No single model wins everything; GPT Image 2 and Wan 3.0 both top specific categories.
• Blind testing: Professionals compare two outputs side by side without knowing which model made either.

Alibaba launched Qwen3.8-Omni-Flash, its first omni-modal model built around agentic tasks, combining audio-video understanding, planning, and tool use in one model while cutting video input costs by about 89%.
Alibaba launched Qwen3.8-Omni-Flash on September 18, its first omni-modal model built for agentic work. It reasons over what it sees and hears, then plans and executes with tools, from auto-editing a vlog to turning a movie into a recap.
It approaches Gemini 3.8 Flash on audio-video benchmarks, gaining 19.5 points on average in agent tests, while video input costs drop about 89%. A 1M-token context lets it scan long videos for key moments using 52% fewer tokens than static analysis.
Key Capabilities:

• Near-Gemini quality: Gains 19.5 points on average against agent benchmarks, closing in on Gemini 3.8 Flash.
• Massive cost cuts: Video input costs drop about 89% compared with Qwen3.5-Omni-Plus.
• 1M-token context: Scans long videos for key moments using 52% fewer tokens than static analysis.