xAI has released Grok 4.6, positioned as a general-purpose model balancing performance, cost, and speed, with significant improvements in handling complex tasks, programming, engineering, and knowledge work. A publicly released model card covers capability evaluations and pre-deployment safety testing, and the team also showcased results on the Artificial Analysis Intelligence Index. Specific scores, API pricing, and availability scope were not disclosed; however, the model is confirmed to be officially launched, with its actual cost-effectiveness awaiting verification through official documentation and independent benchmarks.

The Alibaba Qwen team has open-sourced the weights of the Qwen3.8-2.4T-A95B model, marking the first time weights for a Qwen-Max level model have been made public. The model adopts a Mixture-of-Experts (MoE) architecture with 2.4 trillion parameters, activating 95 billion parameters per token. It natively supports a context length of 256K, extendable up to 1M. Among its 512 experts, 10 are activated per token plus one shared expert, and it employs multi-step token prediction training to enhance end-to-end performance in programming, office tasks, scientific research, and long-horizon Agent tasks.

Anthropic has announced the introduction of invisible, machine-readable watermarks into text generated by Claude, aiming to meet transparency requirements for AI-generated content. The watermark signal persists through copy-paste and minor editing, though the exact implementation remains undisclosed—likely involving statistical patterns formed by word choice. Current information suggests that rewriting or text normalization may still remove these markers, meaning watermarks are better suited for auxiliary identification and溯源 rather than standalone proof of authenticity; detection tools, applicable models, and rollout scope have not yet been disclosed.

Google has added 14 new external service connections to the Gemini app, enabling users to invoke authorized applications within a unified conversation to assist with planning, content creation, and task management. Official examples include reading active deals from Zoho Bigin and summarizing meeting notes from Granola, indicating integration coverage in CRM and knowledge work scenarios. Gemini also announced that new app connections based on MCP will roll out gradually over the coming weeks, though the full list of apps, regional availability, plan restrictions, and third-party data permission rules remain undisclosed.

Anthropic has updated the Claude Chrome experience, allowing the browser sidebar to run Cowork sessions consistent with desktop, web, and mobile versions, enabling users to start tasks in a tab and continue them across devices. Session history, Skills, and Connectors are now available within the browser environment. The feature is currently rolling out to Max and Team users, with Pro users to follow. In a demo scenario, Claude reviewed multiple invoice tabs and wrote amounts into a Google Sheets monthly report, while the team also warned about potential hidden instruction attacks via web pages.

Swedish AI software company Lovable has completed a $400 million Series C funding round, bringing its post-money valuation to $13.3 billion—nearly doubling from eight months prior. The company's annualized revenue is approaching $600 million, with clients including nearly two-thirds of Fortune 500 enterprises, signaling growing demand for natural language generation applications and AI software extending from individual developers to large organizations. Details such as lead investors, funding terms, use of proceeds, and profitability were not disclosed, so the key confirmed developments remain the scale of financing, valuation growth, revenue level, and enterprise customer penetration.

Mistral plans to build up to 1 gigawatt of AI computing infrastructure in Europe by 2030 to expand local model training and inference capacity and strengthen European control over autonomous AI infrastructure. The company has begun seeking long-term customer commitments to secure demand for future capacity planning and funding. Details such as data center locations, phased deployment timelines, chip suppliers, construction costs, or signed customers have not been disclosed, meaning the project remains in the strategic expansion and commercial commitment gathering phase, with actual scale dependent on future contracts and infrastructure progress.

OlmoEarth Studio has introduced on-demand export of base model embeddings, allowing users to select region, month, encoder, spatial resolution, and Sentinel-1 or Sentinel-2 imagery, with results provided as int8 Cloud-Optimized GeoTIFFs. These embeddings support similarity search, few-shot segmentation, change detection, and PCA exploration. Notably, a linear classifier trained on 192-dimensional features using just 60 labeled pixels achieved an F1 score of 0.84 in identifying mangroves, water bodies, and other land cover types.

Google Research has introduced a knowledge profiling framework and the WikiProfile benchmark, categorizing factual knowledge into five states to distinguish between facts encoded versus accessible. The study shows that state-of-the-art models like Gemini 3 and GPT-5 already encode 95% to 98% of tested facts, yet still fail to recall 26% to 34% during direct answer generation. Chain-of-thought prompting can recover 40% to 65% of encoded but inaccessible facts, but only improves unencoded facts by 5% to 15%, suggesting that factuality optimization should shift focus from knowledge expansion to enhancing knowledge retrieval capabilities.

Microsoft Research has released the MindTopo benchmark to evaluate multimodal large models' understanding of five topological relationships: continuity, separation, order, enclosure, and knotting. Test results show that models perform significantly better in static recognition of single scenes than in interactive planning. Even when correctly perceiving initial scenes, models often lose target relationships during multi-step operations, adopt locally rational but globally ineffective actions, or violate environmental constraints. While image generation can sometimes aid single-frame judgment, video reasoning often alters topological structures, making reliable long-horizon planning difficult.