Google and Meta came for Claude

hi guys 👋,

Before we get into it, this newsletter just crossed 13,000 subscribers, which is wild. Thank you to every single one of you for reading.

This week I put Top Carousels inside your AI, so you can pull viral carousels into a chat and remake them right there (more on that at the end).

Google and Meta both shipped models aimed straight at Claude, so the first half of this issue is about whether either one is worth switching for.

There's also a 3-model Claude Code setup worth copying, ChatGPT feeling broken, and an honest look at the free Adobe clones everyone is sharing.

Google has been really quiet this year. Its last Pro model, Gemini 3.1 Pro, came out back in February and set record benchmark scores. Gemini 3.5 Pro was supposed to follow at Google's conference in May, then got delayed over coding problems.

Since then, Google has been teasing Argon for weeks.

Now Gemini 4 Argon is finally out for everyone, and the benchmarks are actually looking pretty good ⬇️

The coolest part is that it makes things up far less often than other models (15% vs 51% for GPT-6 Astra). But after testing it against Opus 5.5, I honestly didn't notice that much of a difference.

My take: Opus 5.5 still feels a bit smarter and nicer to work with, so unless you're already paying for Gemini, I wouldn't switch.

Zuck has been cooking in the dark, and Meta just dropped a new AI model called Muse Wonder.

According to Meta, its top version beats Claude Fable 5.1, GPT-6 Astra and Opus 5.5 on multiple benchmarks (the standard tests AI companies use to compare their models).

Oh, does it?

Because from early testing, it looks like Meta benchmaxxed the whole thing, which means they tuned it to ace the tests rather than to be great at real work. Just look at these examples. There's no way this is better than the top models lol.

To be fair, Alexandr Wang, Meta's AI prodigy, said they didn't specifically train it for 3D games. In my own testing it's not too bad, but it definitely doesn't feel as good as Opus 5.5.

My prediction: Meta has so much money that it'll probably be very generous with usage limits, so the subscription might actually be worth it.

Here's a Claude Code trick I've realised not many people know: run a 3-model team instead of stuffing everything into one expensive model.

Sonnet does the work, hands the boring errands to Haiku, and only rings Opus when it's about to make a big call or gets stuck.

Anthropic calls this the advisor strategy, and in its own tests, the cheapest model more than doubled its score on a research test once it could ask Opus for help.

So you get way better work than the cheap models manage alone, without burning through your limits like you would running Opus on everything.

Update Claude Code, then start it with Sonnet as the main model and Opus as the advisor:

To put the small background jobs on Haiku, add model: haiku to your subagent files.

ChatGPT has been painfully slow for me this week, to the point where it’s genuinely hard to use for real work.

Every single prompt, no matter how simple, was taking me over 5 minutes. So I cancelled my ChatGPT subscription and doubled down on Claude.

Life’s too short.

One developer got so fed up with Adobe's cancellation fees that he used Claude Opus 5.5 to rebuild the whole suite. It's called ArtCraft, and it's a free, open-source set of Adobe look-alikes for Mac, Windows and Linux.

Here's what's actually usable, based on my own testing:

→ PdfCraft (Acrobat): good enough for everyday forms, redaction and organising pages.

→ VectorCraft (Illustrator): fine for logos and icons, but it can't reliably open real Illustrator files.

→ PhotoCraft (Photoshop): fine for basic edits and opens most Photoshop files, but it's glitchy and slows your computer down.

→ FilmCraft, LightCraft, EffectCraft and DesignCraft: okay for simple edits at best, and mostly demos for now.

REA has been going viral alongside these apps. It's a free GitHub tool that lets your AI figure out how pretty much any app works, even without its source code.

Top Carousels now plugs straight into ChatGPT and Claude (via MCP). Ask it for viral carousels, swipe through the slides right there in the chat, and hit Recreate to remake one for your niche, all without leaving the conversation.

It's last week's Krabby Patty formula (find an outlier outside your niche, then rebuild it with AI) without switching tabs.

Setup takes about 2 minutes: copy one address, paste it into your AI's settings, sign in. Step by step in here

→ "Show me the top AI carousels this month."

→ "Find the best-performing fitness carousels and show me the slides."

→ "Show me my saved carousels." (yes, it can see your bookmarks)

→ Paste any topcarousels.com link: "Recreate this for my carousel about AI agents."

Using it inside Claude

Here's what happens when you recreate it

Claude/ChatGPT actually looks at the slides, so your version keeps the layout, fonts and colours that made the original work, with your topic and your words.

It’s out now and you start for free here :)

That's it for this week.

See you next Sunday,