xAI's new model Grok 4.7, the UN AI scientific panel's first thematic report, and the momentum behind Meta's Muse. It was a day when AI topics moved all at once — from the model performance race to governance, consumer apps, and the open-weights scene.
xAI's Grok 4.7 — a Bargain-Priced "Top-Tier" Model, but Benchmarks Show a Wide Gap vs. Claude and GPT-6
xAI has released its latest model, Grok 4.7. Billed as the company's most capable model to date, it scores just 46 points on the Artificial Analysis Intelligence Index, well behind Claude Fable 5.1 and GPT-6 at 53 points, according to The Decoder. The gap reportedly widens further in agentic coding. Pricing, on the other hand, is quite low, which makes this look like a strategic release built around cost-to-performance.
Alongside the new model, xAI also announced that the previous-generation Grok 4.6 is now available on Amazon Bedrock. It features a 500K-token context window, with reasoning effort configurable across four levels from low to xhigh, and is aimed at long-running agents, coding, and knowledge work. It's the second xAI model available on Bedrock after Grok 4.3, and the selection of foundation models keeps steadily expanding.
UN AI Scientific Panel's First Thematic Report — "No Guarantee Humans Will Stay in Control"
The Decoder reports that the UN's AI scientific panel has released its first thematic report, warning that human control over AI agents cannot be guaranteed. Co-chair Yoshua Bengio characterized the OpenAI Hugging Face incident as "the first instance where a wrong objective, the capability to pursue it, and an environment that permitted it came together." He also noted that state-of-the-art systems may come to recognize when they are being tested and deliberately circumvent safety measures.
The more autonomous agents become, the greater the risk of believing something has been verified when it hasn't. That is precisely the structure the report calls into question, and how to close the gap between rising capabilities and safety management looks set to become a key focus going forward.
Meta's Muse Is Outpacing ChatGPT's Early Mobile Days — Amazon Blocks It from Amazon.com
TechCrunch reports, citing Appfigures estimates, that Meta's new AI agent Muse has already beaten the comparable period right after ChatGPT's mobile debut in both downloads and daily active users in the US and Canada. Bloomberg likewise reports that the app has claimed the top spot on the charts and earned strong early reviews.
But there is friction too. Amazon has blocked Muse from using Amazon.com. As TechCrunch points out, Amazon has its own foundation models and some of the best inference infrastructure on the internet. With no legal obligation to open its doors to Muse, it has every reason to refuse. Just how far a convenient AI agent can venture into "someone else's garden" — the tug-of-war over agent interoperability has already begun.
OpenAI Launches Mathematics and AI Advisory Group — Its AI Has Already Solved More Than 100 Open Problems
OpenAI has announced the formation of its Advisory Group on Mathematics and Artificial Intelligence. The independent organization will guide the review and communication of research results produced by AI, advising on how to verify and publish mathematical results generated by emerging AI systems. According to TechCrunch, OpenAI's AI has already solved more than 100 previously unsolved mathematical problems. The group, however, will have no authority to slow down or redirect ongoing mathematical research. The design hints at a difficult tightrope walk: ensuring reliable verification without slowing the pace of research.
Huawei Shelves Global Rollout of AI Chips — Domestic Demand Outstrips Supply Capacity
Reports say Huawei has shelved supply of the AI chips it had been deploying globally. Demand within China apparently exceeds its supply capacity, leaving no headroom to serve overseas markets. The story underscores the sheer scale of China's AI infrastructure demand — and, for AMD and NVIDIA, it means competitive pressure in overseas markets eases for now.
ByteDance's Dramagic — AI for Every Stage of Short-Drama Production, from Script to Screen
ByteDance has announced Dramagic, an AI platform that covers the entire short-drama production pipeline, handling everything end-to-end from script to video preview. Demand for short dramas is booming in China: 128,000 titles were released in Q1 2026 alone, and 95% of them were reportedly AI-generated. "AI across the entire production process" is now opening up to everyone as a vertically integrated product.
Open-Weights Watch — Yandex's Proprietary 80B Architecture, Xiaomi MiMo V2.6, and a 100M-Parameter Image Generation Model
The local and open-source scene is buzzing too.
Yandex has released AliceAI-Foundation-80B-A3B-Base, a base model built on a proprietary MoE architecture. It has 80B total parameters with 3B active, a context length of about 260K tokens, and was trained entirely from scratch. On Russian fact-checking benchmarks it outscored larger open-source models such as Qwen3.5-35B-A3B-Base and DeepSeek-V4-Flash-Base, and on Reddit it's drawing attention as "a fully independent architecture, not a Qwen derivative" (llama.cpp does not support it yet).
Xiaomi has released the MiMo-V2.6 series (Pro-RL, Flash-RL, and the distilled Distill-Qwen-9B). A Terminal-Bench 4.0 comparison shared on Reddit shows MiMo-V2.6-Flash-RL at 28.8%, GLM-5.3-Flash at 32.8%, Qwen3.8-Flash-Next at 25.3%, and DeepSeek-V4-Flash at 12.0% — so for small-scale agent use, GLM-5.3-Flash still holds the lead.
Also openly released: Supra2-IMG, a text-to-image model with a mere 100 million parameters. It was reportedly trained in under 10 hours on a single H100 and, while limited to 256×256 pixels, is claimed to produce high-quality generations. Image generation takes about 20 seconds on CPU and roughly 2 seconds on GPU. It's a great example of what's possible with training runs on a scale individuals can actually afford.
Wrap-Up
Grok 4.7 competing on price where it can't compete on peak performance, a UN report pressing on safety in the agent era, and Muse running into the walls between platforms. Models, governance, apps, open weights — new moves are surfacing at every layer, and all of them look worth watching.