AI Weekly Update - March 6, 2026

Another week, another wave of AI announcements. Here’s what’s actually moving the needle.

OpenAI: GPT-5.4 (Codex 5.4)

OpenAI dropped what they’re calling the biggest update since GPT-4. The headline: native computer use mode - the model can actually interact with your desktop, not just generate text.

Key details:

• 1M token context window - yes, one million
• Financial plugins for Microsoft Excel and Google Sheets built-in
• Combines coding, reasoning, and computer use in one model
• Benchmarks show it outperforming human professionals by 83%
• Available in Thinking and Pro variants

If this works as advertised, the “AI that uses your computer” era is here.

Google: Agent-Ready Workspace

Google made Gmail, Drive, and Docs agent-ready for OpenClaw integration. That’s a big deal if you’re building AI agents that need to touch your productivity stack.

Nano Banana 2 dropped late February - faster image generation + combines text rendering and web search capabilities. Available across Gemini app, Search, AI Studio, and Cloud.

NotebookLM keeps shipping useful features:

• Cinematic Video Overviews - turns your notes into AI-generated videos
• 10 customizable infographic styles
• Full PPTX export (editable decks)
• Zillow integration for home-buying guidance
• Personal Intelligence feature in testing

Cursor: The $2B Coding Assistant

Cursor’s agent update landed, and the revenue numbers are wild - annualized revenue doubled to $2B+. They’re rolling out a new agentic coding tool and launched their first in-house LLM called “Composer” promising 4x speed boost.

The coding tool wars are getting serious.

Qwen3.5: The Open-Weight Disrupter

Alibaba’s Qwen3.5 dropped with a lineup that matters:

• Flagship: 397B params, 17B active (MoE) - Apache 2.0 license
• Medium: Runs on consumer GPU (32GB VRAM), beats old 235B model
• Small: 0.8B to 9B - built for edge, mobile, IoT

Key capabilities:

• Multimodal (text, image, video) - early-fusion, not bolted on
• 1M token context
• 201 languages
• Native thinking mode with <think> reasoning chain
• Built-in tool use + function calling

The kicker: Qwen3.5-Flash API is ~70% cheaper than GPT-5 at $0.10/M input tokens. And it’s all open-weight - run locally, zero per-token cost.

OpenClaw

Reached GitHub’s most-starred status. The agent framework space keeps heating up.

---

That’s the week. The theme: AI agents are becoming real, open-weight models are getting usable, and the pricing war is starting.