The Most Eventful Week of 2026

If last week was busy, this one is downright historic. OpenAI killed its flagship video generation product. A data leak revealed Anthropic's next model. And Shopify launched direct selling inside ChatGPT, Gemini, and Copilot. Meanwhile, NVIDIA unveiled its next GPU architecture and robotics AI raised over a billion in a single week.

As usual, I'm doing the sorting: here's what happened and what it concretely changes for your business.

The Major News This Week

Sora shuts down — March 24: It's official — OpenAI is killing Sora after just 6 months. The numbers are brutal: downloads had plunged 75% from peak, the app was burning roughly $15 million per day in inference costs for a total revenue of... $2.1 million. Disney pulled its $1 billion investment deal. OpenAI is redirecting compute toward robotics and a new project codenamed "Spud." The lesson is clear: a technically impressive product isn't necessarily a viable one. Before investing in an AI tool, always ask yourself: who pays, how much, and why?

Anthropic "Mythos" leaked — March 26: A data leak revealed that Anthropic is internally testing a new model called "Mythos." The interesting part: it introduces a "Capybara" tier that sits above Opus — the most powerful model until now. Described internally as "a step change in capabilities," it shows record scores in coding, academic reasoning, and cybersecurity. Anthropic confirmed the model's existence while acknowledging it poses "unprecedented" cybersecurity risks. If you're building AI workflows, prepare for a major quality leap. And above all, stay flexible on model choice.

Shopify Agentic Storefronts — March 25: This might be the most impactful business news of the week. Shopify merchants can now sell directly inside ChatGPT, Google AI Mode, Gemini, and Microsoft Copilot. Users discover a product through an AI conversation, click, and complete the purchase on the merchant's store. OpenAI takes a 4% commission on ChatGPT sales. This is the beginning of conversational commerce at scale. If you're in e-commerce, activate this immediately — a new acquisition channel was just born.

OpenAI acquires Astral — March 19: OpenAI acquired Astral, the creators of uv, Ruff, and Ty — the most popular Python tools in the developer ecosystem. The team joins Codex, which now exceeds 2 million weekly active users (3x since January). The signal is clear: OpenAI is investing heavily in developer tools. If you use uv or Ruff daily, expect deeper integration with the OpenAI ecosystem.

New Models

Mistral Small 4 — Hybrid MoE architecture: Mistral releases an impressive model under Apache 2.0 license. Mixture of Experts architecture with 128 experts (4 active per token), 119 billion total parameters but only 6 billion active. Result: excellent performance in chat, code, and agents at a fraction of the cost. For self-hosted use cases in Europe (hello GDPR), this is the new go-to. Open-source, performant, lightweight.

NVIDIA Nemotron 3 + Cosmos 3 — GTC 2026: NVIDIA used its annual conference to launch several open model families. Nemotron 3 for multimodal understanding, Isaac GR00T N1.7 for robotics, and Cosmos 3 for "world models." Jensen Huang's message is clear: foundational models must be open for the hardware + software ecosystem to work.

Xiaomi "Hunter Alpha" — The Chinese surprise: A mystery model appeared anonymously on OpenRouter and created buzz in the AI community. Everyone thought it was DeepSeek. It was actually Xiaomi. A new credible competitor in the LLM race, and further proof that China is multiplying its top-tier AI players.

Gemini 3.1 Flash Live — Google goes real-time: Low-latency version of Gemini for real-time multimodal interactions. Google also launches Lyria 3 Pro, its most advanced music generation model, available to Gemini subscribers and via the API.

GPT-5.4 Mini & Nano — OpenAI democratizes: Lighter, more economical versions of GPT-5.4. The trend is clear: the best models are becoming accessible to every budget.

AI Tools to Know

Claude Computer Use — The agent that controls your screen: Claude can now point, click, scroll, and navigate your applications. Available for Pro and Max subscribers in Cowork and Claude Code. Concretely, you can ask Claude to fill out a form, navigate a website, or perform actions in software that has no API. It's one step closer to truly autonomous AI agents. Not yet 100% reliable on complex tasks, but the direction is clear.

Adobe Firefly Custom Models — Your brand, your AI: In public beta, Adobe now lets you train Firefly models on your own creative assets. The result: an image generator that respects your visual identity (style, colors, characters). For marketing teams producing content at scale, this is a game changer. No more generic AI images that don't look like your brand.

NVIDIA Agent Toolkit — The enterprise stack for AI agents: Presented at GTC 2026, it's an open-source platform for developing autonomous enterprise AI agents. Partnerships announced with Adobe, Atlassian, Salesforce, SAP, and ServiceNow. If you're developing AI agents for clients or internally, this is the reference stack taking shape.

Google Gemini Canvas — Search becomes a workspace: Gemini Canvas transforms Google Search AI Mode into an interactive canvas for all US users. It's no longer a list of links — it's a workspace where you interact with results. Google also launched an import tool that lets you transfer conversations from rival AI apps. The message is clear: Google wants ChatGPT users back.

Seedance 2.0 in CapCut — AI video generation for everyone: ByteDance integrates Dreamina Seedance 2.0 directly into CapCut, the world's most popular video editor. AI video generation is no longer reserved for early adopters — it's arriving in a tool millions of creators already use daily.

The Numbers That Matter

Anthropic is considering an IPO as early as October 2026 according to Bloomberg. The race to go public with OpenAI is on.

OpenAI raises an additional $10 billion, bringing the total beyond $120 billion. Staggering amounts — funding tomorrow's AI infrastructure.

AMI Labs, founded by Yann LeCun (former Meta AI chief), raises $1.03 billion in seed funding — the largest seed round in European history. Backed by Bezos, NVIDIA, Samsung, and Temasek. The goal: build "world models" based on the JEPA architecture.

Robotics AI enters the mega-round era: $1.2 billion raised in a single week across Mind Robotics ($500M), Rhoda AI ($450M), Sunday ($165M, new unicorn), and Oxa ($103M). AI is no longer just software — it's moving to the physical world.

NVIDIA announces Vera Rubin, its next GPU architecture after Blackwell, with 3-4x improvement in AI compute density. Jensen Huang projects $1 trillion in orders for Blackwell and Vera Rubin through 2027.

On the market side, US startup funding slows to ~$13 billion in March after the records set in January/February. Europe, meanwhile, hits its highest point of the year.

My Take — What I'm Actually Doing With This

After digesting all these announcements, here are my pragmatic recommendations:

Sora's death is a lesson for everyone → Before adopting an AI tool, question the business model. $15M/day in costs for $2.1M in total revenue is a ratio that kills any product. Prioritize tools with a clear economic model and real traction.

Shopify Agentic Storefronts = new acquisition channel → If you're in e-commerce, activate it now. Conversational commerce is no longer a futuristic concept — it's an active sales channel inside ChatGPT with 200M+ users. 4% commission is the price of a new acquisition channel.

Prepare for the Mythos leap → The Anthropic leak suggests an imminent quality jump. If you're building AI workflows, keep your architecture flexible on model choice. Today's best model won't be the same in 3 months.

Mistral Small 4 for European self-hosting → Apache 2.0, 6B active parameters, solid performance. For companies with GDPR constraints, this is the new standard. It runs on reasonable hardware and doesn't require cutting-edge GPUs.

Watch Claude Computer Use closely → Desktop automation via AI isn't 100% reliable yet, but the trajectory is clear. Start identifying repetitive tasks in your workflows that could be automated by a screen-controlling agent.

See You Next Week

Second edition of this weekly AI watch. Every week, I break down the releases that matter with a practical angle: no hype, no buzzwords, just what you can use starting Monday morning.

If a topic interests you or you want me to deep-dive into a specific tool, let me know. See you next week.