Top Stories
Anthropic Launches Claude for Word, Targeting Legal and Enterprise Markets
Anthropic released Claude for Word in public beta on April 10, bringing a native AI sidebar directly into Microsoft Word on Mac and Windows. The initial pitch is aimed squarely at lawyers and enterprise document teams: the tool reads complex multi-section contracts, processes comment threads, and edits clauses while preserving formatting, numbering, and styles.
The more interesting detail is that Claude for Word connects to Claude for Excel and Claude for PowerPoint, creating a single conversation thread that spans all three open documents simultaneously. Users can ask Claude to flag data inconsistencies between a Word report and its backing Excel model, or align language across a presentation and its source document. That kind of cross-application AI session has been promised by plenty of startups but rarely delivered inside a tool with Microsoft Office’s installed base.
The legal industry is watching closely — and with reason. When Anthropic released its legal plugin for Claude Cowork back in February, Thomson Reuters, RELX, and Wolters Kluwer collectively lost $285 billion in market value in a single trading session. Claude for Word deepens the same strategic bet: AI capabilities deployed where billable hours are highest and manual review is most expensive.
Source: The Next Web, Artificial Lawyer, Let’s Data Science
Google Gemma 4 Ships Under Apache 2.0 — And It’s Not Close
Google DeepMind released Gemma 4 on April 2, and the benchmark improvements over Gemma 3 are large enough to look like typos. AIME 2026 math scores jump from 20.8% to 89.2%. LiveCodeBench coding leaps from 29.1% to 80.0%. GPQA science climbs from 42.4% to 84.3%. The 31B dense variant currently ranks third on the Arena AI open model leaderboard.
The release includes four variants — a 2B, 4B, 26B MoE, and 31B dense — all supporting text, vision, and audio natively across 140+ languages. The smallest models run on smartphones and Raspberry Pi hardware. The 26B MoE variant fits on consumer GPUs with 256K context. Every variant ships under the Apache 2.0 license, meaning unrestricted commercial use, modification, and redistribution without royalties.
This is a meaningful shift from Google’s previous approach. Gemma 3 used a more restrictive license. Apache 2.0 puts Gemma 4 on the same legal footing as Linux or PostgreSQL. For companies that want frontier-adjacent capability without depending on an API provider, this is the strongest open-weight offering available. The combination of multimodal input, long context, and edge deployment makes it genuinely useful for production workloads that the bigger closed models can’t touch.
Source: Google Blog, Engadget, Google DeepMind
Sora Dies in Two Weeks — Disney Found Out an Hour Before Everyone Else
OpenAI’s text-to-video app Sora will shut down on April 26. The API follows in September. The numbers tell the story: user count peaked around one million and dropped below 500,000, while the service burned an estimated $1 million per day in compute costs. A Wall Street Journal investigation confirmed what was obvious from the metrics — Sora was a money pit with no viable path to profitability.
The Disney fallout makes it worse. Disney had committed $1 billion to a Sora partnership and was given less than an hour’s notice before the public announcement. The media giant subsequently pulled out of its planned $1 billion stake in OpenAI entirely. Burning your biggest media partner on your way to an IPO is a choice.
For the broader AI video market, Sora’s failure is instructive. Generating a few seconds of impressive video turned out to be much easier than building a product people actually use regularly at a cost anyone is willing to pay. The gap between “technically impressive demo” and “sustainable business” is wider for video generation than for most AI applications, and no one has bridged it yet.
Source: TechCrunch, Variety, OpenAI Help Center
Quick Hits
-
AI weapons race escalates without rules: A New York Times investigation found U.S. military autonomous drone programs are lagging China’s capabilities. China recently tested about 200 autonomous vehicles simultaneously in a drone-swarm exercise. Congress’s March 26 primer on lethal autonomous weapon systems acknowledged there’s still no agreed international definition of what these systems are — let alone rules governing them. The Secretary-General’s 2026 target for a legally binding treaty looks increasingly aspirational. Prism News, Creati.ai
-
Google bakes NotebookLM into Gemini: Notebooks rolled out this week in the Gemini app, giving users personal knowledge bases that sync bidirectionally with NotebookLM. Upload PDFs, documents, URLs, and YouTube videos through Gemini’s side panel, and they appear in NotebookLM automatically — and vice versa. Video overviews and infographics created in NotebookLM become accessible from Gemini. Available now for AI Ultra, Pro, and Plus subscribers on web, with mobile and free tiers coming. Google Blog, Dataconomy
-
Open-source model flood continues: Beyond Gemma 4, this past week saw Alibaba ship Qwen 3.6-Plus with a 1-million-token context window and agentic coding capabilities under a hybrid MoE architecture. Microsoft released three MAI foundational models — MAI-Transcribe-1 (speech-to-text, 50% lower GPU cost than alternatives), MAI-Voice-1 (60 seconds of speech in under one second), and MAI-Image-2 (text-to-image, #3 on Arena.ai). Nvidia’s Nemotron 3 Super shipped with 120B parameters, 12B active, designed for multi-agent systems. The barrier to entry keeps dropping. Caixin Global, SiliconANGLE
-
US AI chip export policy hitting bottlenecks: Bloomberg reports that licensing delays and staff shortages are undermining America’s push to control AI chip exports. The gap between policy ambition and administrative capacity is widening just as Chinese labs demonstrate they can train frontier models on non-Nvidia hardware. Bloomberg
Worth Watching
The week’s model releases tell a clear story about where the market is splitting. Anthropic is going deep into enterprise integration — Claude for Word, Excel, and PowerPoint form a unified AI session across the Microsoft Office suite, targeting the highest-value workflows where switching costs are enormous. Google is going wide with open weights — Gemma 4 under Apache 2.0 makes frontier-adjacent capability available to anyone with a GPU and a use case. Both strategies are coherent. Both can work. They don’t compete as directly as the leaderboard rankings suggest.
Meanwhile, the Sora shutdown is a useful corrective to the narrative that every AI capability automatically becomes a product. OpenAI’s most technically impressive consumer offering failed not because the technology was bad but because the economics didn’t work and the use case was unclear. That’s worth remembering the next time someone announces an AI feature and assumes adoption will follow.