Weekly AI Roundup (3/28): Sora Shut Down, Apple's New Siri Delayed Again, MCP Nears 100M Installs

Sora is disappearing. The new Siri is delayed once more. Meanwhile, the infrastructure for AI agents keeps solidifying — the last week of March 2026 made the AI industry’s “focus and prioritize” moment crystal clear.

Here are five stories you need to know this week.

1. OpenAI Shuts Down the Sora App and API

On March 24, OpenAI announced it will discontinue the Sora AI video generation service, including its iOS app, API, and website.

Just months after its public launch late last year, Sora peaked at 3.3 million downloads but had dropped to 1.1 million by February. The primary reason for the shutdown is the enormous inference cost. Revenue couldn’t keep pace, so OpenAI is redirecting resources toward world-simulation research for robotics.

Sora 2 video generation will remain available within ChatGPT’s paid plans.

Key takeaway: AI video generation is technically feasible, but building a sustainable business around it turned out to be far harder than expected.

Sources: TechCrunch / Bloomberg

2. Apple’s New Siri Delayed Again — Pushed to iOS 26.5 Beta

The overhauled Siri, which was set to be rebuilt entirely on LLMs, has been delayed once more.

Originally slated for iOS 26.4, Bloomberg reports that issues with query processing accuracy and response speed have pushed it back to an iOS 26.5 beta release. Additionally, reports indicate that iOS 27 will open Siri up to third-party AI assistants beyond ChatGPT.

Key takeaway: Even a company the size of Apple struggles to bring LLMs to production quality. Their commitment to polish is admirable, but managing user expectations remains a challenge.

Sources: Bloomberg / Ubergizmo

3. MCP (Model Context Protocol) Surpasses 97 Million Installs

Anthropic’s Model Context Protocol (MCP) crossed 97 million installs in March.

MCP is a standard protocol that enables AI models to connect with external tools and data sources. Every major AI provider now ships MCP-compatible tools, marking its evolution from an experimental spec to core agent infrastructure.

Key takeaway: As AI agents move toward real-world deployment, having a common language for tool connectivity is a big deal. For developers, MCP support is quickly becoming a de facto requirement.

Source: Crescendo AI

4. March Was a Frontier Model Bonanza — 4 Models in Just 23 Days

March 2026 may be the most densely packed month in AI history.

GPT-5.4 (3/5): 1.05M token context window, Standard/Thinking/Pro variants
Gemini 3.1 Flash-Lite (3/3): Ultra-low-cost at $0.25 per million input tokens
Grok 4.20 (3/22): Enhanced real-time web search
Qwen 3.5 Small (3/1): Multimodal text/image/video, 0.8B-9B parameters

Four frontier models dropped in just 23 days. The competitive cycle among AI labs has compressed from months to weeks.

Key takeaway: Developers have more options than ever, but choosing which model to adopt — and when — is getting increasingly difficult.

Sources: devFlokers / Build Fast with AI

5. Alibaba’s “Wukong” — An Enterprise Multi-Agent Platform

Alibaba, China’s largest e-commerce and cloud computing company, has unveiled “Wukong,” an enterprise-grade AI platform.

Multiple AI agents collaborate to autonomously handle business tasks like document editing, approval workflows, and research. It integrates with messaging tools like Slack and Teams, as well as existing enterprise systems.

Key takeaway: We’re shifting from “one AI agent that does everything” to “multiple agents working as a team.” The way enterprises adopt AI is fundamentally changing.

Source: AI and News

The Bottom Line

This week’s theme: focus and prioritize. OpenAI let go of Sora, Apple chose quality over speed, and meanwhile agent foundations like MCP and Wukong are steadily expanding.

AI’s evolution isn’t slowing down, but in 2026, the decisions about what to stop building may matter more than what to start.

Related reading:

This article was generated by AI and reviewed by humans.