Hi {{first_name}}!
What a week in AI. GPT-5.4 dropped, Anthropic told the Pentagon no and ended up at #1 in the App Store as a result, and there are a handful of meaningful updates across the board worth knowing about, from new tools to a security warning your team should hear.
This week we're covering:
Ten AI updates from across the space, including new models, big moves, and one security warning worth paying attention to
A Cowork workflow I built that turns a single meeting transcript into six automatic outputs without me touching anything
An idea I've been sitting on for a while that I'd genuinely love your input on
Let's get into it.
New and Noteworthy
OpenAI launches GPT-5.4 OpenAI released GPT-5.4 this week, and this looks to be one of the more meaningful updates for real business work in a while. Early access testers are calling it the best all-purpose model right now. The focus is not just better chat. It is stronger performance on spreadsheets, presentations, documents, and computer use. OpenAI says GPT-5.4 is 18% less likely than GPT-5.2 to produce responses containing any errors, and it posted major gains on spreadsheet modeling and presentation tasks. The bigger signal here: AI is becoming more useful for actual deliverables, not just brainstorming. GPT-5.4 Thinking is available for Plus subscribers ($20/month) and above while GPT-5.4 Pro is reserved for the $200/month ChatGPT Pro and Enterprise plans.
ChatGPT for Excel is here: OpenAI also launched ChatGPT for Excel in beta, which brings GPT-5.4 directly into workbooks. You can ask it to build or update models, run scenarios, explain formulas, and generate outputs based on cells and formulas. This is exactly the kind of update that could save operators, finance teams, and business owners real time, especially as these tools get more embedded in the software people already use every day. OpenAI also announced new financial data integrations alongside the Excel launch.
Claude hits #1 in the App Store after Anthropic says no to the Pentagon: Anthropic CEO Dario Amodei held firm on two points: no mass surveillance of Americans, no fully autonomous weapons. The Pentagon wouldn't budge. The government directed federal agencies to stop using Anthropic products. Anthropic plans to fight it in court. What happened next? Daily sign-ups broke records all week, free users are up more than 60% since January, and paid subscribers have more than doubled this year. Meanwhile, ChatGPT uninstalls surged 295%. The bigger takeaway: consumers noticed a company holding a red line, and they responded. A reminder that sometimes the best growth strategy is just deciding what you won't do.
Notion’s AI now works across your whole workflow: Notion expanded its Custom Agents so they can work across Slack, Notion Mail, Notion Calendar, and MCP-connected tools. That is worth watching because it points to a broader shift: AI is moving beyond the side chat window and into the actual systems where work happens. For teams trying to reduce friction, this is where things get interesting. The value is not just asking better questions. It is having AI show up closer to the flow of work.
Microsoft 365 Copilot gets GPT-5.3 Instant: Microsoft announced that GPT-5.3 Instant is now rolling out inside Microsoft 365 Copilot and Copilot Studio. They're positioning it as an upgrade for everyday use, with more accurate responses, stronger writing, and fewer unnecessary disclaimers or refusals. That matters because a lot of teams will not adopt AI through brand-new tools. They'll feel it through the assistants already showing up inside Microsoft, Google, and the rest of their existing stack.
Google releases Gemini 3.1 Flash-Lite, its fastest and cheapest model yet: Built for speed and volume, Flash-Lite runs 2.5x faster than the previous Flash model at $0.25 per million input tokens. It's designed for high-volume tasks like data processing and customer communications where you need fast, cost-effective AI at scale. If your team is using AI in large quantities, this one's worth a look.
Anthropic expands Cowork with enterprise plugins: Anthropic added 13 new MCP connectors and a new wave of pre-built plugin bundles covering HR, design, engineering, finance, and operations. New connectors include Google Workspace, DocuSign, and FactSet, meaning your team can now have Claude working directly inside the tools you already run your business on. If you're on a Team or Enterprise plan and haven't explored plugins yet, this is a good week to start.
Anthropic updates skill-creator so you can test and improve your AI agent skills: This one is more relevant if you're already building with Claude Skills, but the concept matters for any team investing in AI workflows. Anthropic added the ability to write evals (essentially automated tests) that check whether a skill is doing what you expect. You can now benchmark performance, catch regressions when models update, and tune skill descriptions so they trigger at the right time. The short version: your AI workflows can now be tested like software, not just vibed-checked.
Apple's new MacBook Pro is built from the ground up for AI: The updated 14- and 16-inch models with M5 Pro and M5 Max chips are less of a hardware refresh and more of a statement about where Apple is heading with on-device AI. LLM processing is up to 4x faster than the M4 generation, and a Neural Accelerator is now built into each GPU core, dramatically speeding up tasks like image generation and on-device model training. If your team is running AI tools locally or processing large datasets without wanting to send everything to the cloud, this machine handles it without breaking a sweat.
Microsoft warns about malicious AI browser extensions: This is a useful reminder that AI adoption also needs guardrails. Microsoft reported that malicious browser extensions impersonating AI assistant tools collected LLM chat histories and browsing data from platforms like ChatGPT and DeepSeek. Those extensions reportedly reached about 900,000 installs across more than 20,000 enterprise tenants. If your team is experimenting with AI, this is a good moment to review browser extension policies and be more thoughtful about what tools are getting installed.
Tool Spotlight: Claude Cowork
I've mentioned Claude a lot lately, and honestly, that's not by accident. There's a reason it's sitting at the top of the App Store right now. Over the past few months, it has become one of the most capable and reliable LLMs out there, and it keeps getting better. So yes, you're going to keep hearing about it from me.
But there's one amazing new feature that was rolled out within the last few weeks, and I want to give it the proper attention it deserves. It's not flashy but every time I use it, I think about how much time it saves me. That's Cowork.
Quick overview for anyone still getting up to speed: Cowork is Anthropic's autonomous desktop agent. It's accessible via the Claude app (not via their web based app) and it is a program that you install on your Mac. Unlike most AI tools that live in a browser tab, Cowork runs locally on your computer, which means it can actually see your files, interact with your desktop apps and take action inside the software you use every day. The difference from regular Claude chat is that instead of responding to one prompt at a time, Cowork plans and actually executes multi-step tasks on your behalf, accesses files directly on your computer, connects to your existing tools, can use your Chrome browser, and can run on a schedule, all without you having to babysit it.
Here's what that looks like in practice. I drop a raw meeting transcript into Cowork. No additional prompt. Just the transcript.
From there, it takes over:
Generates a structured meeting summary
Saves it to Google Drive as a branded document
Drafts a follow-up email and drops it into my Gmail drafts, already populated with the contact's email address and subject line
Updates my CRM if it was a prospect
Adds the meeting notes to the respective opportunity card in our CRM
Pulls out any action items for me and my team and creates tasks directly in ClickUp, my project management system, with all details, appropriate priority flags, relevant transcript context, and due dates
One input. Six outputs. I don't touch anything. And as a safety precaution, you have the option to review and approve each step before Cowork proceeds.
So how is this possible? The workflow is built on top of something called Skills. Think of Skills as reusable instruction sets you build once inside Claude. Each one packages a specific set of steps, your preferences, your tools, your output format, so Claude knows exactly what to do and how to do it without you repeating yourself every time. This workflow is made up of multiple Skills working together in sequence, each one handing off to the next automatically.
The setup takes some time upfront. But once it's built, every meeting feeds itself into my systems automatically.
I've put together a full guide covering Claude Cowork in depth , how to set up Skills, how to connect your tools, privacy settings to configure before you start, and how to think about building your first workflow. If you're interested in a copy let me know!
Considering an In-Person Mastermind
Not another networking event. Not a lecture. A working mastermind experience.
Something's been on my mind lately and I'd genuinely love your input.
I'm craving more real, in-person time with smart, curious operators who want to use AI to build better businesses. If you're running a business and serious about using AI to operate smarter, not just experiment with it, this is for you.
Small group. Unique venue. Great food. Real conversation. Laptops open.
This would be 8–12 people max. Small enough that everyone's voice matters and real relationships actually form.
I’d give a quick presentation on AI updates then we’d go deep on 2–3 focused topics and actually build together. One topic I of course want to dive into right away: autonomous agents and how to deploy them in a practical and responsible way.
This could take the form of monthly half-day sessions, a full-day hackathon-style experience, or even a quarterly multi-day retreat somewhere worth traveling to.
You'd leave with real workflows built, new thinking on where to take your business next, and a small group of peers you can actually call on.
What excites me most is the human aspect. The depth of conversation you only get when thoughtful people are together solving real problems. AI is moving fast. The advantage won't just come from tools. It'll come from the quality of collaboration around them.
I've already had a handful of conversations with people who are in. Now I'm curious how many of you feel the same way.
What format would make it worth your time? Hit reply and let me know. I'm still shaping this and want to build it around the people who show up.
That's it for this week! If you're interested in the in depth Claude Cowork guide I put together, hit reply and I'll send it over.
See you next week,
Julien
PS: If you know someone that would value the information I share each week please forward this to them, or send them a link to subscribe at www.ampra.ai/join-our-newsletter.