Week in AI: Alibaba Goes Proprietary, Gemma 4 Beats the Giants, and GPT-4o Dies at 23 Months

Week in Review: March 30 - April 4, 2026

This was the week the open-source AI landscape cracked down the middle. Alibaba — the most prolific open-source AI lab outside of Meta — released three new models and made all of them proprietary. Within 48 hours, Google fired back with Gemma 4, proving that open-weight models can still beat proprietary giants on key benchmarks. Between those two bookend stories, OpenAI killed GPT-4o, Microsoft shipped AI agents powered by Claude, and MCP became the undisputed infrastructure standard for AI tool use.

Here is everything that happened.

1. Alibaba Releases Qwen3.6-Plus — All Closed-Source

April 2, 2026 — Alibaba released Qwen3.6-Plus, a frontier-class model engineered for agentic coding with a one-million-token context window. The headline capability: it leads all models on Terminal-Bench 2.0, the agentic terminal coding benchmark, scoring 61.6 versus Claude Opus 4.5's 59.3. On SWE-bench Verified, it scores 78.8 — second only to Claude's 80.9.

The real story is not the benchmarks. It is the licensing. Qwen3.6-Plus is available exclusively through Alibaba Cloud's API, the Qwen App, and approved third-party tools. None of Alibaba's three April releases are open-source. This represents a complete strategic reversal from a company that open-sourced every previous Qwen generation and had over 100 million downloads of Qwen3.

The shift to proprietary is driven by economics: analysts estimate Qwen3.6-Plus cost approximately $800 million to train, and only 11 percent of previous Qwen deployments ran on Alibaba Cloud. The company was spending hundreds of millions to benefit AWS and Google Cloud. Goldman Sachs published a note titled "The End of the Free Model Era" predicting other Chinese labs will follow within twelve months.

Our analysis: The Great AI Closing: Alibaba Goes Proprietary and the Open-Source AI Dream Starts to Die

2. Google Gemma 4 Launches — Open-Source Fights Back

April 3, 2026 — Google released Gemma 4, a family of open-weight models that immediately challenged the narrative that proprietary models are necessarily superior. The 31-billion-parameter Dense variant landed at number three on the Arena AI text leaderboard, beating proprietary models backed by billions in infrastructure. The 26B MoE variant hit number six while running on a single 80GB GPU.

Most impressively, the 2-billion-parameter "Effective" model runs on a Raspberry Pi — answering complex questions offline with no internet connection. Google is proving that the open-source ecosystem can produce genuinely competitive models when a well-funded lab commits to the strategy.

The timing — one day after Alibaba's proprietary pivot — was almost certainly deliberate. Google's Gemma team lead published a blog post reaffirming Google's commitment to open models.

Our analysis: Google Gemma 4 Changes Everything: When Open-Source AI Models Start Beating the Giants

3. OpenAI Retires GPT-4o After Less Than Two Years

April 3, 2026 — OpenAI officially retired GPT-4o from all plans, less than 23 months after its revolutionary debut. The model that introduced real-time multimodal interaction and was once considered the industry's gold standard is now legacy software.

The retirement highlights the brutal upgrade cycle in AI: cutting-edge becomes obsolete in months, not years. For enterprises that integrated GPT-4o into production workflows, the deprecation forced a scramble to migrate to GPT-5.4 — which has different capabilities, different pricing, and different behavior that breaks existing prompts and integrations.

Our analysis: The End of GPT-4o: What OpenAI's Model Retirement Tells Us About the AI Industry's Brutal Upgrade Cycle

4. Microsoft Ships Claude-Powered Copilot Cowork Agents

April 4, 2026 — Microsoft announced Copilot Cowork, a new enterprise AI agent system powered by Anthropic's Claude Agent SDK. This is significant for two reasons: Microsoft is using a competitor's AI (Claude) alongside its own OpenAI partnership, and the system represents the first major enterprise deployment of truly autonomous AI agents from a hyperscaler.

Copilot Cowork agents can autonomously manage project workflows, coordinate across Microsoft 365 applications, and make decisions within policy guardrails — a step beyond the simple copilot assistants that preceded them. The Claude Agent SDK integration provides the reasoning backbone while Microsoft handles the enterprise integration layer.

Our coverage: Microsoft's Claude-Powered Copilot Cowork Signals the Enterprise Agent Era

5. MCP Crosses 97 Million Installs

March 30, 2026 — Anthropic's Model Context Protocol crossed 97 million installs in March, cementing its position as the universal standard for AI tool connectivity. Every major AI provider now ships MCP-compatible tooling, and the protocol has become the default mechanism for agents to connect to external tools and data sources.

The number matters because MCP is becoming infrastructure — as fundamental to AI applications as HTTP is to web applications. Developers building agentic systems that do not support MCP are increasingly building for an ecosystem of one.

Our coverage: MCP Crosses 97 Million Installs: How One Protocol Became the Universal AI Connector

6. HSBC Eyes 20,000 AI-Driven Job Cuts

April 1, 2026 — HSBC is reportedly considering eliminating up to 20,000 positions through AI-driven automation, focusing on middle-office functions, compliance processing, and routine customer service. If executed, it would be the largest single AI-related workforce reduction by a financial institution.

The timing coincides with OpenAI hitting $25 billion in annualized revenue and Meta unveiling custom AI inference chips — signals that AI infrastructure is maturing to the point where large-scale enterprise deployment is economically viable.

Our coverage: HSBC Weighs 20,000 AI-Driven Job Cuts as OpenAI Hits $25B Revenue

7. AI Commoditization Wave Accelerates

April 3, 2026 — The open-source tool ecosystem continued its explosive growth. OpenClaw, the open-source alternative to Claude Code, crossed two million active users. Meta is reportedly considering significant layoffs in departments where AI agents have already replaced human workflows. The gap between proprietary and open tools continues to collapse even as the gap between proprietary and open models widens.

Our coverage: AI Commoditization Wave: OpenClaw Explodes, Meta Eyes Mass Layoffs

Additional Coverage This Week

Monday: Deploy a Remote MCP Server with Streamable HTTP on Cloudflare Workers — Step-by-step tutorial
Tuesday: The Economics of AI: When the Math Doesn't Work — Deep analysis of AI spending vs returns
Wednesday: Pusher Channels MCP Server on the Official Registry — Our open-source contribution
Thursday: How AI Will Replace Customs Brokers — HAR Series

Week Ahead: April 6-11, 2026

NVIDIA Earnings Preview. NVIDIA reports Q1 FY2027 earnings on April 9. The market is watching for data center revenue guidance in light of the Vera Rubin platform announcement and increasing competition from custom chips (Meta MTIA, Google TPUv6, Huawei Ascend 950PR). Expectations are for $45-48 billion in quarterly revenue.

DeepSeek V4 Expected. Multiple sources indicate DeepSeek will release V4, a one-trillion-parameter mixture-of-experts model trained on domestic Chinese hardware, within the next ten days. If the model is competitive with Qwen3.6-Plus while remaining open-source, it could partially counterbalance Alibaba's proprietary pivot.

EU AI Act Enforcement Begins. April 7 marks the first enforcement deadline for the EU AI Act's highest-risk categories. Companies deploying AI in critical infrastructure, employment, and law enforcement must demonstrate compliance or face penalties of up to 35 million euros or 7 percent of global revenue.

OpenAI IPO Filing Watch. With OpenAI generating $2 billion per month in revenue, market watchers expect an IPO filing in Q2 2026. Any filing would trigger mandatory financial disclosures that could reshape understanding of the AI industry's actual economics.

The Big Picture

This week crystallized the central tension in AI for 2026: the models are getting more powerful and more expensive, which means the organizations that build them are getting more motivated to monetize rather than share. Alibaba's proprietary pivot is the biggest signal yet that the free model era is ending — but Google's Gemma 4 proves the open-source community is not going quietly.

For developers, the message is clear: build abstraction layers, maintain multi-model capability, and do not assume any single provider's openness is permanent. The AI landscape is bifurcating into a proprietary tier for frontier capabilities and an open tier for everything else. Where the line falls between those tiers will define the next chapter of this industry.

Sources: