I paid for Google’s Gemini Ultra plan ($100/month) to get early access to Gemini Spark, and I came away with a split opinion: this is one of the most capable hosted agents I’ve used, and I still can’t tell most people to buy it yet.

Spark is closer to “a junior operator on call” than anything else I’ve tried. It takes longer-running tasks, uses more compute, and works across your files instead of just chatting about them. The catch is that all of that power stops at the edge of Google’s ecosystem—and at $100/month, that edge is hard to ignore.

Here’s what stood out after hands-on testing.

What Spark does really well

Docs and Sheets workflows are fast and seamless. My first test was simple: can Spark actually do work inside the Google suite? It passed. Creating docs, spinning up spreadsheets, and moving between them was quick and genuinely impressive. If your day is “outline a plan, build a sheet, summarize a thing,” Spark delivers value on day one.

Calendar feels native, not bolted on. Scheduling-style tasks were smooth, which tracks given how deep Calendar already sits in the Gemini experience. It read like a real extension of what Google does well, not a demo feature.

It personalizes using everything Google already knows about you. When I asked it to put together a plan for my toddlers, it already knew how many kids I have and their ages—no restating. That’s equal parts creepy and useful. Creepy because you feel the surveillance-adjacent reality of living inside one company’s ecosystem. Useful because the agent finally stops acting like a goldfish. If you’ve been waiting for AI tools to remember context, this is a real glimpse of it.

Where it gets interesting: longer-running tasks

Auditing my YouTube channel was a standout. I had Spark review my channel and write up what’s working, what isn’t, and what to do next. It’s strong at pulling info from the web, synthesizing it, and handing back something actionable. The remote-browser view—where you watch it work—makes it feel less like magic and more like an operator you can supervise.

The chore automation is the real tell. I set up a recurring workflow to watch my email for rewards-card updates and log new offers to a master sheet. Yes, an app already exists for that. The point isn’t novelty—it’s that Spark collapses the setup cost. There was no integration to wire up, no automation tool to learn. I described the chore in plain English and it just ran. That’s the upside of this whole category: quietly removing dozens of micro-tasks from your week without the usual config tax.

The drawbacks—and why $100 is a hard sell

The integration gaps are oddly glaring. There’s no Google Tasks or Google Keep support. Those are already Google products, and they’re exactly the lightweight inputs that would make an agent feel sticky. Leaving them out makes the ecosystem story feel half-finished.

There’s no real extensibility, and that’s the dealbreaker. Spark doesn’t connect to anything outside Google’s walls. No way to plug in third-party tools or data sources—nothing like MCP (Model Context Protocol), which is fast becoming the standard for letting agents securely reach other systems. For comparison, the hosted agents I’ve used from Anthropic and OpenAI lean hard into exactly this kind of openness, and it’s what makes them useful for real cross-system work. Spark is powerful right up until you try to extend it past Google, and then it hits a wall.

So is Gemini Spark worth $100/month?

For most people, not yet. If your work is genuinely Google-suite-heavy—docs, sheets, calendar, web research—and Spark can reliably offload it, you can probably justify the cost today. For everyone else, it’s still early.

But I want to be clear about which way this leans: I used Spark more than I expected to, and it’s one of the most promising Gemini implementations I’ve touched. The problem isn’t capability. It’s that the price is running ahead of the openness. The day Google adds real extensibility, I’d resubscribe without thinking about it.

What would flip it to an instant yes:

1. Native Tasks and Keep support, at minimum

2. A first-class integration layer (MCP-style extensibility)

3. Third-party tool coverage beyond Google’s core suite

4. A lower tier for light-agent use, so the price matches the current capability

Until then, Spark is a compelling preview of what hosted agents can be—just not the everyday workhorse it’s trying to become.

1. GPT Image 2.0 — The jump I wasn’t expecting

For anything image-related, I’ve been living in Nano Banana and that family of models. Cheap, fast, good enough.

GPT Image 2.0 feels like a different class of tool. I gave it my website and asked for a redesign — one shot, no chain-of-prompts. What came back was genuinely usable. Not “AI slop you then fix,” but something a designer would actually start from.

What stood out:

• Text in images finally works — headlines render cleanly.

• Layout intuition is real — it understood hierarchy without me spelling it out.

• Style consistency across variants — three options felt like a set, not three random generations.

If you’ve been ignoring OpenAI for image work because Nano Banana was “good enough,” this is worth reopening the tab for.

2. Claude Design — Not a designer killer, but a great thinking surface

I don’t think it replaces a designer yet. But as a wireframing and prototyping canvas, it’s solid. Getting rough ideas out of my head and onto something I can click around with — that’s the sweet spot.

Where it struggled: full redesigns were mixed, and credits disappeared fast, which cut my testing short.

Where it shined: quick prototypes and early-stage exploration when I don’t know what I want yet.

Not a stack-changer — but it’s earned a slot.

3. Capacities MCP Server — Quiet, useful, exactly what MCP is good at

Capacities is a note-taking app I really like, and they just shipped their MCP server.

Here’s the pattern I keep finding with MCP: it’s magic when you want to push info from your AI chat into a tool. Capture an idea mid-conversation, drop it into the right object in Capacities, done.

Less flashy than the two above — but that’s kind of the point. It just works, and now my notes app is one less place I have to manually sync.

Key Takeaways

• GPT Image 2.0 — legitimate quality jump over Nano Banana.

• Claude Design — great for wireframes, not full design replacement yet.

• Capacities MCP — small but meaningful upgrade if you’re already in the app.

• MCP keeps proving its value as the glue between AI chats and your real tools.

Full GPT Image 2.0 redesign breakdown (plus prompts) coming to the Stack Snacks YouTube this week. 👉 [youtube.com/@stacksnacks]

P.S. — If you’ve tested GPT Image 2.0 on your own, hit reply and send me the output. Want to see how far this stretches.

Hey Snackers 👋

Anthropic dropped Claude Opus 4.7 last week — an upgrade to their flagship model — and this one’s worth a real look.

The headline: this is a coding model you can actually hand work off to. Not “hand it simple stuff and babysit it” — hand it hard, multi-step, hours-long tasks and walk away. At least, that’s the claim.

But there’s more going on under the hood than just better code. Better eyes, tighter instruction following, and a brand new approach to cybersecurity that’s quietly one of the most interesting things any AI lab has done this year.

Let’s break it down.

The Big Upgrade: Software Engineering

The main event with Opus 4.7 is coding. Anthropic says early testers were able to delegate their hardest, most supervision-heavy tasks to the new model with confidence — something they couldn’t reliably do with Opus 4.6.

What makes it different? Three things:

1. Self-verification. The model checks its own work before reporting back. It catches logical faults during planning and doesn’t just plow ahead hoping for the best.

2. Long-horizon consistency. Multi-step tasks that span many turns — the kind that used to degrade or lose the plot — now stay on track.

3. Benchmark results. State-of-the-art on SWE-bench (the standard software engineering benchmark), including multilingual and multimodal variants. On a 93-task coding benchmark, it improved resolution by 13% over Opus 4.6 — including 4 tasks that no previous Claude model could solve.

One early tester put it well: “Low-effort Opus 4.7 is roughly equivalent to medium-effort Opus 4.6.” Translation: it’s doing more with less thinking time.

4 More Things Worth Knowing

1. 3x Better Vision

Opus 4.7 accepts images up to 2,576 pixels on the long edge (~3.75 megapixels). That’s 3x the resolution of previous Claude models. If you’re using Claude for reading screenshots, extracting data from diagrams, or any computer-use agent work — this is a big deal. Heads up: higher-res images eat more tokens.

2. Instruction Following Got Too Good

This one’s interesting. Opus 4.7 takes instructions much more literally than previous models. Anthropic actually warns that prompts written for older versions might produce unexpected results — because older models would loosely interpret or skip steps. If you’re migrating from 4.6, re-tune your prompts.

3. Better Memory Across Sessions

The model got better at using file system-based memory, meaning it remembers context across multi-session projects. Less “let me re-explain everything” at the start of each chat.

4. Tastier Creative Output

Anthropic says the model produces higher-quality interfaces, slides, and docs. Less “AI-looking,” more polished. This matters if you’re using Claude for anything client-facing.,

Pricing & Availability

• Price: $5/M input tokens, $25/M output tokens (same as Opus 4.6)

• API string: claude-opus-4-7

• Available on: Claude.ai, API, Amazon Bedrock, Google Vertex AIv

• Tokenizer change: Updated — same text may use 1.0–1.35x more tokens than 4.6

Migration note: The tokenizer change means your costs could creep up even at the same per-token price. Anthropic recommends measuring real-world token usage on your workloads before fully switching over.

Also launched yesterday:

• xhigh effort level — a new setting between “high” and “max” for finer control over reasoning depth vs. speed

• Task budgets (public beta) — guide Claude’s token spending across longer runs

• /ultrareview in Claude Code — a dedicated review session that audits your code for bugs and design issues

• Auto mode for Claude Code Max — lets Claude make permission decisions on your behalf for longer, uninterrupted task runs

The Bigger Picture

Three trends converging here:

1. The “unsupervised coding” bar keeps rising. Six months ago, handing an AI a complex multi-file refactor and walking away was a fantasy. Now the top models are genuinely getting there. Opus 4.7’s self-verification isn’t a gimmick — it’s the feature that makes long-running agentic work practical.

2. Capability shaping is the new safety conversation. Forget content filters and guardrails. Anthropic is experimenting with making models structurally less capable in dangerous domains. That’s a fundamentally different approach than “add a warning label.” If it works, expect every lab to follow.

3. The model treadmill is exhausting — but the gaps matter. Yes, there’s a new model every week. But Opus 4.7 vs. 4.6 isn’t incremental — 13% better on hard coding tasks, 3x vision resolution, and a new safety architecture. If you’re building on Claude, this is a real upgrade.

Should You Care?

If you code with Claude daily: Yes. Full stop. The self-verification and long-horizon improvements are real. Start with your hardest tasks and see if the claims hold up. Re-tune your prompts — the instruction following is tighter.

If you’re building AI agents: The vision upgrade (3x resolution) and better memory across sessions make computer-use agents significantly more viable. Worth testing.

If you’re watching the AI safety space: The Project Glasswing / capability shaping approach is a genuinely new move. Agree or disagree with it, but pay attention — this is likely where the industry is heading.

🔗 Anthropic’s official announcement | Project Glasswing | Cyber Verification Program

💬 Hit reply and tell me: are you switching to 4.7 immediately or waiting? I read every response.

Until next time,

John

Here’s the thing nobody’s talking about enough: we’re not just using AI anymore – we’re building with it.

A few months ago, making your own AI agent required Python scripts, API keys, and a tolerance for broken things. Now? You can spin one up in plain English inside tools you’re already paying for.

I’ve been deep in two platforms this month that are making custom AI agents feel... normal. Here’s what I found.

Notion Custom Agents

What it is: Notion launched Custom Agents in February 2026 – autonomous AI teammates that live inside your workspace and respond to triggers (Slack messages, database changes, schedules, emails). Over 21,000 agents have already been created during the beta.

What makes it interesting:

• You describe what you want in plain English – no code

• Agents can pull from your existing Notion pages, databases, and docs as their knowledge base

• They integrate with Slack, Mail, Calendar, Figma, and more

• Scoped permissions mean you control exactly what each agent can access

Pricing: Free during beta (through ~April 2026) on Business and Enterprise plans. After that, it moves to a credits system — $10 per 1,000 Notion Credits.

My take: I built a few agents inside my own workspace – one that monitors my content calendar and drafts newsletter outlines when I add a new topic, and another that answers questions about my project docs. Setup was shockingly fast. You describe what you want, point it at the right databases, and it just... works. The trigger system is where it gets really powerful – I have one that fires whenever a new entry hits my ideas database. What surprised me most is how well it pulls context from existing pages. It’s not just keyword matching; it actually understands the relationships between my docs.

Verdict: If you’re already a Notion power user, this is a no-brainer to try while it’s free. The learning curve is basically zero.

🔗 Try Notion Custom Agents

Hyperagent (by Airtable)

What it is: Hyperagent is Airtable’s dedicated autonomous agent platform – a standalone product built to run complex, hours-long workflows across enterprise data. It gives agents a full computing environment – they can learn new tools and APIs on the fly, chain them together, and produce finished deliverables. Think: hours of research or data work, condensed into minutes.

What makes it interesting:

• Agents run autonomously with access to a real computing environment

• They learn and adapt to new tools and APIs as they go – no pre-built integrations required

• Can chain multiple tools together to complete complex, multi-step workflows

My take: What gets me about Hyperagent is the combination of breadth and depth. The integration list is genuinely impressive – it connects to the tools you’d actually want an agent touching. But the real unlock is chaining. You can wire together a sequence of complex actions and just... let it run. I’ve set it loose on tasks I’d normally block out an hour for and come back to finished work. That’s a different feeling than most automation tools, which still need you hovering. The learning curve is real – this isn’t a no-code play. But if you’re dev or dev-adjacent, that tradeoff is worth it fast.

Verdict: This is the power-user pick. If you need agents that can handle complex, multi-step work across the open web, Hyperagent is worth exploring.

🔗 Check out Hyperagent

What This Means

Both of these tools point in the same direction: AI agents are becoming personal infrastructure.

Notion is betting you’ll build agents inside your existing workspace. Hyperagent is betting you need agents that roam the open web.

Different approaches, same destination – AI that works for you in the background, not just with you in a chat window.

My prediction: by the end of 2026, everyone reading this will have at least one custom agent running. The tools are that accessible now.

Honestly, setting these up scratched the same itch as building automations in Zapier years ago – except now the automations can actually think. They read context, make judgment calls, and adapt. That’s the shift. We went from “if this, then that” to “figure out what to do and do it.” And we’re still in the early innings.

📺 Want to see these in action? I’m doing deeper walkthroughs on the YouTube channel – subscribe to Stack Snacks on YouTube so you don’t miss them.

💬 Hit reply and tell me: have you built any custom agents yet? What are you automating? I read every response.

Until next Tuesday,

John

In this post, I’m walking through three real tools I built from scratch on the Zo platform, each using a different open source model and a single prompt. Practical builds that show you what these models can actually do.

Why I’m Using Zo for This

These open source models – MiniMax, Kimi, GLM – don’t have the same plug-and-play API access as the frontier models. You need somewhere to run them, and that’s where Zo comes in.

[Zo](https://zo.computer) is a single workspace for hosting web apps, running AI models, and building tools. What drew me to it wasn’t just the model hosting – it’s that you can go from prompt to deployed app without stitching together a separate frontend, backend, and hosting stack. Everything lives in one place.

Right now, they’re offering free access to all three models, which made it easy to run real builds without worrying about burning credits. I’m on the basic paid plan, but the free tier gives you enough runway to figure out what these models can actually do.

Build #1: Link-in-Bio Page (MiniMax M2.5)

For the first test, I wanted something simple – a link-in-bio page hosted directly on my Zo space. I gave MiniMax M2.5 a single prompt with all my links (Stack Snacks, newsletter, YouTube, etc.) and told it to build the page.

**How it went:** There was a little bit of funkiness getting it to render initially. The first attempt failed, and I had to restart the chat. But once it got going, it built a clean, colorful link-in-bio page that absolutely gets the job done. It’s not going to win any design awards, but for a one-prompt build? Solid.

I also ran the same prompt through GLM-5 and Kimi K2.5. GLM-5 noticed the existing page and basically cloned it, while Kimi K2.5 was so fast it had the page up before I could even check the preview. The speed on all three models was impressive, even if the link-in-bio use case was too simple to really differentiate them.

**Verdict:** All three models handled this basic task without breaking a sweat. The real differences showed up in the next builds.

Build #2: Credit Card Rewards Calculator (Kimi K2.5)

This is where things got interesting. I wanted to see if Kimi K2.5 could do something more complex: research a brand-new credit card (the Robinhood Platinum), pull all the benefits and perks, and build me an interactive calculator to figure out if the annual fee is actually worth it.

One prompt. That’s it.

**How it went:** Kimi K2.5 went to work immediately. It ran a bunch of web searches, read through multiple articles about the Robinhood Platinum card, compiled all the offers into a markdown file, and then started building the calculator. The whole process took just a few minutes.

The result was a fully interactive tool where I could plug in my estimated spending across categories – travel credits, hotel credits, DoorDash, dining, flights, car rentals, everything else – and see whether the card’s annual fee pencils out based on my actual habits.

After running my numbers, the calculator showed me a net annual value and even compared it against the Amex Gold card. It’s exactly the kind of quick analysis I’d normally spend 30 minutes doing manually in a spreadsheet.

**Verdict:** Kimi K2.5 impressed me the most here. It handled the web research, math, and UI building in one shot. Fast, accurate, and genuinely useful output.

Build #3: AI App Tracker Database (GLM-5)

For the final build, I wanted to test GLM-5 on something with more of a coding challenge – a database-driven app. The prompt was simple: build me a tracker for all the AI apps I’m using, what I’m paying for, and a list of things I want to try. I asked for a table with tags and told it to pre-populate some of the major AI tools.

**How it went:** GLM-5 took a different approach than the other models. It built a proper little API to store the data, then created a front-end page to consume it. It even added authentication, which I hadn’t asked for – a nice touch for anything that might eventually hold real data.

The final result was a clean table interface with tags like “want to try,” “paid for,” and “stopped.” Each entry could be edited inline through an edit panel at the top of the page. You could delete entries, update statuses, and open the actual tools directly from the tracker.

Was it perfect? No – the edit functionality needed a small nudge. But for a single-prompt build that included a backend API, a JSON data store, and a responsive front-end? That’s remarkable.

**Verdict:** GLM-5 is the coding powerhouse of the group. It handled the most technically complex task without breaking a sweat.

My Final Rankings

After a few weeks of using all three models regularly, here’s where I’ve landed:

**Kimi K2.5** is my pick for general-purpose work. It’s fast, handles research tasks well, and produces clean output across a range of use cases. This one has become part of my daily workflow.

**GLM-5** is the one I reach for when I need something more technically complex. If the build involves APIs, databases, or heavier code, GLM-5 handles it with confidence.

**MiniMax M2.5** is solid across the board, but didn’t stand out as strongly as the other two in my testing. Still absolutely worth having in the rotation.

The Bigger Takeaway

Here’s what really changed for me: these models shifted *how* I allocate my AI usage. I’m now using open source models for the everyday stuff – quick tools, simple builds, research tasks – and saving my frontier model tokens (Claude, GPT-4, etc.) for the complex work that truly needs them.

If you’re a solopreneur, creator, or small business owner, this is huge. You can build real, functional tools without burning through expensive API credits or paid subscriptions. The quality gap between open source and frontier models is shrinking fast, and for a lot of daily tasks, it’s already negligible.

Try It Yourself

Head over to [Zo](https://zo.computer) and sign up for free access while it’s still available. Pick a small tool you’ve been meaning to build – a calculator, a tracker, a simple landing page – and throw a single prompt at one of these models. I think you’ll be surprised at what comes back.

If you want more AI tool breakdowns like this, check out the [Stack Snacks newsletter](https://www.stack-snacks.com/) or subscribe to the channel on YouTube.

---

*Which of these three models are you most interested in trying? I’d love to hear what you’d build with them.*

If you’ve been following the AI agent space, you’ve probably seen the buzz around running your own AI computer—setting up Mac Minis, virtual private servers, and dealing with messy configurations. It’s powerful stuff, but it’s not for everyone.

That’s where Zo comes in. It’s a tinkerer’s dream that gives you a lot of that power right out of the box. No server setup. No complex configurations. Just a powerful AI assistant with access to your file system, services, and automations.

Let me walk you through what I built in under 30 minutes.

What Makes Zo Different

At its core, Zo is a chat interface with access to multiple AI models—Gemini, GPT, Claude Sonnet and Opus, plus solid open-source options like GLM 4.7 and the Qwen series. But the real magic happens when you connect it to your actual life: your file system, your calendar, your phone.

The combination of AI + file system + services opens up possibilities that fundamentally change how you interact with these tools.

Four Things I Built (And You Can Too)

1. A Custom Read-It-Later App

We all have articles we want to save and actually read later. Instead of using yet another app, I built my own in Zo. It’s simple: save a link, get access to the full plain text version, and iterate on the design whenever I notice something I want to change.

Want to build one yourself? Zo has templates to get you started. I spun up a basic markdown blog in minutes—just picked a template, named it, and started chatting with that instance as my context. Pure vibe coding.

2. Claude Code Integration

Here’s where it gets interesting. Open a terminal in Zo and you’ve got Claude Code with full file system access. I navigated to my new blog folder, asked Claude to create some sample blog posts, and within a few exchanges had a working blog pulling content from a posts folder.

The speed is addictive. What might take an hour of manual setup happens in a couple of minutes through conversation.

3. Scheduled Tasks That Actually Help

This is where Zoe enters territory that makes it genuinely useful for daily life.

I’ve got two small kids. Doctor’s appointments, school events, daycare coordination—miss one of these and you’re caught off guard at the last second. So I set up an automation that searches my Google Calendar for medical appointment keywords and sends me a weekly heads-up.

I also get my calendar texted to me every morning. A quick summary helps me prepare for the day before I even open my laptop.

One of my favorite features? One-off reminders. Got a piece of mail about a car tax bill? Text yourself a reminder through Zo and it’s handled. No app switching, no remembering to set an alarm.

4. A Phone Number You Can Actually Talk To

This is the part that still feels magical to me: Zo gives you a phone number you can text. You can ask it to kick off automations, remember things, or add notes.

Picture this: you’re driving and have a thought. Fire off a voice-to-text to your Zo number. By the time you get home or to the office, that thought is captured, maybe even acted upon. I’ve used this to create documents, store subscription information, and capture ideas that would otherwise disappear.

The novelty hasn’t worn off. There’s something about texting an AI that knows your context and can actually do things that changes how you approach these tools.

Skills: The Building Blocks

For those unfamiliar, skills are essentially repeatable sets of instructions you can give your AI system. Think of them as templates for common tasks—”convert this markdown to a professionally formatted PDF” with specific details about what “professional” means to you.

The cool part? Skills are portable. If you’re moving from another platform to Zo, your skill files probably work here too. And Zo comes with a library of pre-built skills from the community, Anthropic, and other sources.

I installed a skill for working with MCP servers while recording this video. The ecosystem keeps growing.

Building a Skill in Real-Time

I wanted a content repurposer—something that takes a YouTube transcript and creates both a newsletter version and a LinkedIn version (more professional and proper).

I fed it a transcript, asked for both versions, and within minutes had usable drafts. The LinkedIn version needed tweaks (they always do), but the newsletter version was solid. And here’s the kicker: it automatically saved the blog post into my blog’s posts folder.

Then I saved the whole workflow as a skill. Now I can run “YouTube to Blog” whenever I need it. Just like that, a repeatable process.

Why This Matters

What Zo demonstrates is how powerful it becomes when you tie an AI to three things: a file system, your services, and communication channels.

Each piece individually is useful. Together, they’re transformative.

You can build internal dashboards. Automate reminders based on calendar events. Create content pipelines. Store and retrieve information through text messages. Run code and see results immediately.

And all of it happens through simple conversation. No coding required (though you can go deep if you want to).

The Bottom Line

Zo has become my daily driver. Not because it’s perfect, but because it removes the friction between having an idea and making it happen.

The setup that would require servers and technical configuration with other solutions? It’s just there. The integrations that would take hours to build? They’re templates you can customize in minutes.

If you’ve been curious about AI agents but intimidated by the setup, Zo is worth a look. And if you’re already building with these tools, I’d love to see what you’re making.

That changed this week.

Claude Cowork is here

Anthropic launched Cowork on January 12th, and I haven’t stopped thinking about it since.

The pitch is simple: Claude Code for the rest of your work. All the power of their coding agent — reading files, editing documents, creating new content, automating tasks — but through the familiar chat interface of the Claude Desktop app.

You point Cowork at a folder. You describe what you need. It does the work.

No terminal. No commands. No technical setup.

Simon Willison called it “a general agent disguised as a developer tool” — and that’s exactly right. The capabilities were always there. Anthropic just built a front door that doesn’t require a CS degree to walk through.

- Expanded to all Pro subscribers ($20/month)

- Currently: macOS only, research preview

Here’s what blows my mind: Anthropic built the entire Cowork feature in roughly a week and a half — using Claude Code itself. That’s the kind of recursive improvement that makes you realize we’re in a different era.

I’ve experimented with Claude Code on enough tasks to know this changes things. Work is about to feel different. And now that barrier to entry just dropped dramatically.

Oh, and ChatGPT has ads now

Almost forgot — OpenAI also announced their [advertising approach] this week.

Free tier users will see sponsored content at the bottom of chats. The twist: you can chat with the ads. Ask questions, get details. It’s an interesting format, even if the whole thing feels like an inevitable step toward the Google-ification of AI.

**The good news:** $20/month subscribers won’t see ads. And there’s a new $8/month “Go” tier for people who want expanded features (reasoning models, image creation, file uploads, memory) without the full price tag.

I get it. Running AI infrastructure is expensive. Ads fund access for people who can’t pay. But I’d be lying if I said I wasn’t a little sad to see that clean, ad-free interface change.

## Two different bets

OpenAI is betting on scale and monetization. Anthropic is betting on capability and utility.

Both might work. But only one announcement made me immediately want to build something.

If you’ve been waiting for AI agents to feel usable, Cowork might be your moment. I’ll be experimenting more and sharing what I find.

Until next week,

John

Impressions and Early Experiences

• Embarking on this journey just three days ago, I’ve already witnessed its rapid development, with live updates continually enhancing its capabilities. Sure, it’s still in its alpha stage, which you notice with the need for a paid subscription right from the start—a potential barrier for some. Yet, despite its youthful quirks, the experience has been nothing but promising.

• A standout moment was forwarding an email to my agent, posing a question, and receiving a well-thought-out response. It felt like a peek into the future, one where autonomous agents handle tasks effortlessly. Many apps boast about autonomy, but few deliver like Do Anything.

Key Features and Functionalities

• What truly caught my eye was the app’s coding prowess. It integrates a Claude Code instance when needed, writing code like a pro. Its Markdown file creation keeps my plans organized and easily exportable. The folder system is intuitive, enabling seamless uploading and document management.

• While I haven’t delved into all its connectors—understandably so in its alpha stage—the idea of linking personal data offers thrilling possibilities. Of course, we need to tread carefully to prevent it from becoming too overbearing, but the excitement is palpable.

Personalized Interaction and Future Potential

• Here’s a neat touch: you can name your agent. This personalization elevates the experience, making email interactions with your named assistant feel extraordinary. It’s like glimpsing the future just around the corner.

• The Do Anything app is a powerhouse of potential. I’ve yet to see anything exactly like it, and if there are products out there that match up, I’d be eager to know. For now, Do Anything delivers a compelling experience, nudging us closer to a world where digital assistants seamlessly integrate into our lives.

1) ChatGPT can now see your medical records

OpenAI launched ChatGPT Health — a dedicated tab where you connect your health data and get personalized answers.

Connects to: Apple Health, Peloton, Function, plus your actual medical records via b.well.

The pitch: Ask things like “How’s my cholesterol trending?” or “Help me prep for my doctor appointment” and get answers based on your real data — not generic advice.

Why I think this is a big deal: 40 million people already ask ChatGPT health questions daily. Most of those answers are generic because the AI doesn’t know anything about you. Now it can. For people without easy access to healthcare — or who have questions they’re embarrassed to ask a doctor — this is a genuine unlock: a non-judgmental AI that knows your actual numbers and history.

The caveat: AI confidently states wrong things sometimes. This isn’t a replacement for medical care. But as a first pass, a way to understand results, or prep for appointments? Useful.

Availability: US only. Waitlist open now.

2) Gemini is now baked into Gmail

Google announced Gmail is “entering the Gemini era.” Translation: Gemini 3 is now deeply integrated into your inbox.

What’s new:

• AI Overviews: Ask your inbox questions in plain English (“What did the client say about pricing?”)

• AI Inbox tab: A beta view that filters noise and shows what matters

• Auto to-dos: Identifies action items from emails (pay this bill, schedule that appointment)

• Writing tools: Help Me Write, Suggested Replies, Proofread

What you need to know:

• The question-answering feature requires a Google AI Pro or Ultra subscription

• You can opt out of Gemini, but only by disabling all smart features (including package tracking and calendar integration)

• US + English only for now

Email is overwhelming. AI summarization and smart filtering makes sense here. The question is whether you’re comfortable with the trade-off.

3) Amp Code: Claude Code, but with model choice

Amp is a new coding agent that works like Claude Code — but with a key difference: you pick the model.

Supports: Claude Opus 4.5, Gemini 3, GPT-5

The hook: $10/day free tier (ad-supported).

Works in: Terminal, VS Code, Cursor, Windsurf, JetBrains, Neovim

Early feedback is positive. Users say it feels “more agentic” — you can let subagents run and they reliably produce working code. Works on repos already built with Claude Code.

Quick comparison:

• Amp: Best for teams or anyone who wants model flexibility

• Claude Code: Best for terminal-first solo devs

• Cursor: Best for a traditional IDE feel

I haven’t tested it yet, but the free tier with Opus 4.5 access is hard to ignore. Planning to give it a spin this week.

Bottom line

These aren’t hype launches. They’re AI embedding into things you already use every day: health tracking, email, coding.

ChatGPT Health could be transformative for people underserved by traditional healthcare. Gmail AI feels inevitable (and opt-out-able… sort of). Amp Code gives you options in a space that’s been Anthropic-only.

Worth paying attention to.

John

It’s the most important shift in the AI landscape this year.

Gemini 3 Pro tops the LMSYS Arena. Claude Opus 4.5 dominates coding tasks. GPT-5.2 wins on versatility. Grok 4.1 undercuts everyone on price. Each claim is true. None of them helps you decide.

Here’s why that’s actually great news—and what it means for your wallet in 2026.

Competition finally caught up

Two years ago, OpenAI had a comfortable lead. One year ago, Claude started closing the gap. Today? Four companies ship frontier models within weeks of each other, and users can barely tell them apart in blind tests.

This compression at the top benefits you directly:

• Prices are falling fast. Anthropic cut Claude Opus pricing by 66%. Grok charges $0.20 per million tokens. Google offers generous free tiers.

• Switching costs dropped. Models are increasingly interchangeable for most tasks.

• Innovation pressure is intense. No one can coast.

What I actually use (after testing everything)

Gemini 3 Pro for anything integrated with my Google workflow. The ecosystem connection matters more than benchmark differences.

Claude Opus 4.5 for code. The 66% price cut made this obvious. Opus handles complex, multi-file reasoning better than anything else I’ve tested.

Nano Banna for image generation. Google’s tool has gotten so good and it is baked right into Gemini.

GPT-5.2 when I need reliability across varied tasks. It’s the Honda Accord of AI — never disappointing.

Grok 4.1 for bulk tasks where I’m cost-sensitive. At that price point, the quality-per-dollar ratio is unmatched.

The 2026 spending strategy

Forget finding the “best” model. Build a stack:

Primary tool: Match your ecosystem. Google user? Gemini. Microsoft user? Copilot. Apple user? Wait and see what they ship.

Specialty tool: Pick one for your highest-value use case. For me, that’s Opus for code.

Budget tool: Grok or whatever’s cheapest when you need volume.

Total cost: $20-40/month covers most professionals. That’s less than your streaming subscriptions for tools that actually make you money.

The real question for 2026

The models are good enough. All of them. The question isn’t “which is best?” anymore.

The question is: “Am I using AI for the right tasks?”

Most people still underutilize these tools. They’re paying for subscriptions they barely touch, or using AI for trivial tasks while doing hard work manually.

The winners in 2026 won’t be the people who picked the “right” model. They’ll be the people who integrated AI deeply into workflows that matter.

Stop optimizing your model choice. Start optimizing your usage.

ChatGPT Apps

You can now run apps directly inside ChatGPT. This means you can call up specific applications during your conversations. For example, you can bring up Canva while …

Read more

Notion just dropped version 3.0, and the headline feature is customized AI agents. The full agent product isn’t live yet, but the improvements to Notion AI are already worth paying attention to.

Here’s what’s new and what it means for how you work.

Custom Instructions: Train Your AI

The standout feature is custom instructions. You write detailed instructio…

Read more

The Superlist app's AI features for meeting notes and task management offer an intriguing solution for users looking to enhance their productivity through technology. After spending some time exploring these capabilities, it's evident that the app shines in certain areas while needing improvement in others.

Meeting Notes

One of the standout features of th…

Read more

GPT‑5 Codex is a GPT‑5 variant optimized for agentic coding workflows. It is positioned as a code-first assistant and (at least for now) isn’t available as part of ChatGPT Codex.

• What it is: A GPT‑5 build tuned for autonomous code reading, planning, and multi‑file edits

• Why it matters: Faster refacto…

Read more

“Hey there, Happy Tuesday! 👋 In this snack-sized issue, we are going over the GPT 5 release and some of our earlier thoughts.

GPT-5 Is Out

The day we've all been waiting for is here. I can't remember the last time I was so excited about a product launch that I added it to my calendar to watch the livestream. The consensus is that GPT-5 operates different…

Read more

Kimi‑K2 in 90 Seconds

• Moonshot AI—an Alibaba‑backed Beijing startup founded by Yang Zhilin—released Kimi‑K2 on July 11. It's a 1‑trillion‑parameter Mixture‑of‑Experts model with 32B active parameters per token step, available in both base and instruct versions with fully open weights.

• Thanks to its open nature, developers are already deploying it across m…

Read more

ChatGPT's new Agent Mode is an exciting development that merges the capabilities of their existing operator and deep research products. For those unfamiliar, the operator function allows the AI to take control of a web browser to complete tasks, while deep research involves conducting extensive web searches and providing detailed reports. Agent Mode com…

Read more

1️⃣ Build Your Own AI App With Claude

Claude just rolled out “Artifacts,” a new feature that lets you turn a plain-text prompt into an interactive mini-app. Flip on the “Create AI-powered…

Read more