OpenAI and Anthropic kinda took the week off when it came to big releases.

Everyone else didn't. Google, Microsoft, Amazon, Replit, and Manus all shipped fresh aI features that quietly reshape how you actually work.

A document creator finally inside Gemini. Outlook that runs your inbox while you sleep. A persistent cloud computer that doesn't reset. Real-time voice agents that don't break when callers interrupt.

Calm before the storm? Probably. Google IO is a couple weeks out and the big labs are cooking.

But this week's "small" updates are the kind that sneak into your workflow Monday and make you wonder how you lived without them. If your team slept on this one, y'all are prolly about to fall behind.

Let's get into it.

1. Gemini Finally Made Files 📄

Google Gemini just got a feature that should have shipped six months ago.

You can now generate complete downloadable files straight from a Gemini chat. Google Docs, Sheets, Slides, PDF, Word, Excel, CSV, LaTeX, markdown, RTF, and plain text.

Anthropic was first to market with this. OpenAI followed. Gemini was on an island.

Files save to your device or export straight to Google Drive. Rolling out now to all Gemini users at no additional cost.

Try This

Re-run your last three Gemini chats that ended with "now I have to copy this into a doc," and export straight to Drive.

Cuts your end-of-chat busywork to zero and keeps the formatting your team actually expects.

2. Outlook Goes Agentic 📥

Outlook used to be where you worked. Now it's where Copilot works for you.

That's Microsoft's pitch, word for word. They just shipped Copilot agent mode through Outlook's Frontier early access program.

What changes: Copilot runs continuously in the background to manage your inbox and calendar. Always on. Email triage, drafting follow-ups for unreplied threads, creating inbox rules, resolving meeting conflicts, rebooking rooms, and blocking focus time.

Rollout covers Outlook for Windows, web, iOS, and Android. Microsoft 365 Copilot license required. EU users, sorry. Not at launch.

The catch: it's preview. Bulk actions may be incomplete.

Try This

If your org is in Frontier early access, flip on agent mode tomorrow and let it triage your inbox for a full day.

Half your unreplied threads come back with drafts waiting before you even open Outlook.

3. Amazon Quick Lands On Your Desktop 🖥️

Amazon wants in on the always-on assistant race.

The new native desktop app for Amazon Quick (formerly QuickSuite) runs continuously in the background, accesses local files without uploading, sends OS-level notifications, and automates browser-based tasks.

The unlock is a personal knowledge graph that learns your people, projects, and relationships across every connected tool.

Integrates with Google Workspace, Microsoft 365, Salesforce, Slack, Teams, and Zoom. Preview on Mac and Windows in US East only. No AWS account required. Free and Plus plans.

Feels like Amazon's version of Copilot at the OS level mixed with OpenClaw Lite. Whether it gets traction is the open question.

Try This

If you live inside the AWS ecosystem, sign up with a personal account this weekend and let it learn your workflow.

Once the knowledge graph builds, you stop re-explaining context every time you open a new chat.

4. Replit Ships Slides For Vibe Coders 🎨

Another week, another AI slide tool.

This one's from Replit. Start from scratch, build templates, or import existing PowerPoint files. The AI restyles them via a PowerPoint-to-code converter. Export to PowerPoint, PDF, or Google Slides.

Access is still vague. Replit's changelog says it's live but plan requirements aren't confirmed. The site says "start creating for free."

The interesting part is the converter. Bring an ugly branded deck in, let it beautify the design, keep your brand assets intact. Pulls Replit beyond dev tooling into productivity for non-coders.

Try This

Take your ugliest current deck, the one with mismatched fonts and the slide you're embarrassed to share, and drop it into Replit Slides this weekend.

A clean restyle means one less Canva subscription and one less designer favor to call in.

5. DeepSeek v4 Drops With A Shrug 🤷

DeepSeek v4 was supposed to change the world.

It didn't.

It came as a preview with two variations: v4 Pro and v4 Flash. Both are mixture-of-expert models with a 1 million token context window and up to 384k max output. v4 Pro is 1.6 trillion parameters. v4 Flash is 284 billion. Open source under MIT, weights on Hugging Face, both support thinking and non-thinking modes.

v4 Pro pricing: $1.74 per million input tokens, $3.48 per million output. v4 Flash needs about 160GB of RAM to self-host.

Top-five open source model on benchmarks. Not even the best Chinese open source model right now. The big labs and the US government have accused Chinese companies like DeepSeek of model distillation.

Try This

Skip it unless you're running high-volume workloads where cost-per-token outweighs everything else.

For legitimate US businesses, the compliance headache cancels out the cheaper input pricing.

6. Manus Cloud Computer Stays Awake 🌥️

Every Manus chat used to start from a blank slate.

Not anymore. Manus just launched Cloud Computer, a persistent always-on virtual machine that runs bots, Python scripts, or software 24/7.

Files, installed tools, and context now carry across sessions. Access is via SSH or the Manus web dashboard terminal. No coding required.

Available globally to all paid users on web and mobile. Three plans: Basic (Python scripts), Standard (active websites and APIs), Advanced (team databases).

The catch: it can't access your local machine files. Fully isolated. For some workflows, that's exactly the point. Direct competitor to Perplexity Computer, just earlier and running in Manus's own cloud.

Try This

If you're already a Manus subscriber, spin up a cloud computer this weekend and point one bot at a Slack channel for support triage.

That's a 24/7 worker handling the repetitive triage that used to eat your team's mornings.

7. Copilot Studio Voice Agents Go Live 📞

Traditional IVR breaks the moment a customer interrupts.

Microsoft just shipped the fix. Real-time voice agents in Copilot Studio are now generally available, launching first through Dynamics 365 contact center.

These are speech-to-speech agents optimized for low latency and natural interruption. Callers can speak normally, change direction mid-call, and the agent adapts in real time instead of following rigid scripts.

Calls escalate to a human agent with full conversation context carried forward. No repeating yourself.

Available now in Dynamics 365 contact center for North America. Outside NA must allow cross-geo processing since the model is hosted in NA only. EU, not getting it.

The IT angle: same governance, permissions, and compliance controls applied to voice as text. Over 80% of Fortune 500 companies already have active agents in Copilot Studio.

Try This

Pull recordings of your three most common inbound call types and test them against a Copilot voice agent in a sandbox.

Faster first-call resolution and lower call-handling costs without ripping out your existing stack.