greetings people,

Apologies for going MIA last week.

I’ve been pouring blood, sweat and tears into “The AI Course”.

Early-bird spots open on Monday. Make sure you’re on the waitlist so you don’t miss out (how good are these eye rolling urgency tactics).

buckle up this newsletter is a JUICY read :)

📌 TL;DR

• Fable 5 → Anthropic's first public model above Opus, the best thing ever shipped. then the US government forced them to switch it off worldwide, 3 days later.

• Apple's new Siri → finally a proper conversational assistant, and its real edge is seeing your texts, photos and calendar, context no other AI can touch.

• The Chinese models → GLM-5.2, Kimi K2.7 Code and MiniMax M3 all landed, and they're dirt cheap. but they're still not better than Opus/GPT-5.5, and I don't even use them. here's when you should (and shouldn't).

• Reve 2.0 → a 70-person lab landed #2 on the image leaderboard, ahead of Google's Nano Banana 2, off the back of a clever "layout-first" approach.

• Kickbacks → Claude Code has ads now, and you get paid to watch the spinner. I tested it both as a consumer and an advertiser.

Fable 5: the best model ever made, gone in 72 hours

Ok, this is the big one.

On June 9th Anthropic launched Claude Fable 5, their first public "Mythos-class" model, a whole tier above Opus 4.8.

Mythos-class basically means they took their most dangerous, most capable internal model (Claude Mythos) and made a version safe enough to hand to the public.

They priced it at $10 / $50 per million tokens (in / out), double Opus 4.8, their next best model.

And it was pretty bloody good.

The crew at every.to got a full week with it before launch, and their write-up matched everything else I saw:

• Stupidly good at hard engineering. On their senior-engineer test it scored 91/100, basically a real senior dev. The best before it were Opus 4.8 at 63 and GPT-5.5 at 62.

• You can leave it running for hours. Stripe pointed it at a 50M-line codebase migration and it cleared in a day what would've taken a team two months.

• It's got real taste. It sweats the small details and adds things you didn't think to ask for, instead of doing the bare minimum.

It's really a power-user tool, though. If you already orchestrate agents it'll do things you haven't seen before; if your setup's basic, you honestly won't feel the difference.

And it's slow and hungry, happily burning a million tokens on a single task, so it's overkill for quick everyday jobs. The consensus, theirs and mine: a beast for big, long-running jobs where you hand it a real plan and let it run.

If you want to watch it work, the video running it through 5 one-shot UI/UX tests is a great watch here.

Then on June 12th, it was gone.

The US government issued an export-control order and forced Anthropic to switch Fable 5 and Mythos 5 off for every customer on the planet.

Anthropic is fighting it hard, arguing it's a narrow capability GPT-5.5 already has, and that pulling a model used by hundreds of millions over one flaw would freeze new releases across the whole industry.

Their reasoning was national security: a model this good at reading code and spotting security holes is dangerous in the wrong hands.

Honestly, I didn't really get to test out its full capabilities. I gave it one task - building out a diagram for a podcast I'm recording - and it did an exceptional job. then I got busy… and now it's banned.

So that was that. I’m genuinely looking forward to using it properly when it comes back, and I reckon it will, once Anthropic and the government sort some kind of arrangement out.

Stay tuned.

Apple's new Siri: late and safe, but it sees what nothing else can

At WWDC, Apple finally showed off a proper next-gen Siri.

It's super conversational now, lives in its own standalone app, and your history syncs across iPhone, iPad and Mac. More importantly, it can actually read and do things inside your apps - Messages, Photos, Calendar, Reminders - and it cites the sources so you can tap to check.

Siri can see your texts, your photos, your calendar - the personal context that ChatGPT, Claude and Gemini simply can't touch. That's the real moat here, and the one thing none of them can copy.

It's also deliberately cautious. It'll add the concert to your calendar, but it won't go off and buy the tickets for you (unlike Google's much more aggressive agent demo).

But it's not all rosy. Third-party apps are clunky: you have to name the app, and you can't set a default yet. The proper on-device version only runs on the iPhone Air and iPhone 17 Pro, so even the iPhone 16 misses out.

And it's not actually out yet - developer preview now, full launch around fall 2026 with iOS 27, English first, no EU or China at launch.

(Under the hood it runs on Apple's own models, trained on Google's Gemini and partly running on Google's cloud, on a reported ~$1B/year deal)

Honestly, I’m rooting for them on this one. I'm a big Apple bloke myself, all my stuff is apple, and if they can finally turn Siri into the Jarvis-style assistant we were promised all those years ago… that'd be brilliant for all the little personal-assistant bits of your day.

The Chinese open-source models are incredible now, but should you actually use them?

Funny timing: the same week the US switched off its best model, China handed us three brilliant open-weight ones. Well well well…

The three on the table this week:

• GLM-5.2 (Z.ai) → a usable 1-million-token context window, and on a paid Z.ai coding plan it plugs straight into Claude Code with a quick settings change - no custom workarounds, since Z.ai built an endpoint for exactly that (just know it chews through your plan's usage fast).

• Kimi K2.7 Code (Moonshot) → the open-weight coding workhorse, built for long agentic jobs and unusually token-efficient (Moonshot reckons ~30% fewer "thinking" tokens, which adds up fast on long runs).

• MiniMax M3 → open weights just dropped on Hugging Face, and it's right up there with the best - independent testing has it #2 among open-weight models, just behind GLM-5.2.

Where they win, hands down, is price, and the gap is mad.

A job that costs you about $1 on the cheapest Chinese model (MiniMax M3, at its launch pricing) runs about $21 on Opus 4.8 and about $25 on GPT-5.5 - roughly 20-25x more per word for the US frontier models. (At MiniMax's normal pricing the gap's closer to 10x, which is still ridiculous.)

It's amazing how good these have got, and being open-weight gives us some strong options. That said, I personally don't use any of them, and here's my honest read on when you should.

If you run a business with really sensitive data that can't go through someone else's model (cannabis, peptides, fintech, legal) then being able to own and self-host the model matters enormously, and that's exactly when these come into play.

They're also much cheaper, so if you're on a tight budget running agents like OpenClaw or Hermes, they're a solid option. And you don't even have to host them yourself, you can route them through something like OpenRouter and still pay a fraction of what the top US models cost.

But for me, they're still not better than the state-of-the-art.

Opus 4.8 and GPT-5.5 are better all-round. The Chinese models might top them on a specific benchmark, but not as an all-round model. I just want the best of the best, I’m happy to pay $200-300 a month for it, and I don't need the data-privacy angle. So personally there's no reason to switch.

And honestly, for most people right now, I don't think there is either.

Reve 2.0: the image model built around control (and it just hit #2)

A tiny independent lab just gatecrashed the image wars. Reve AI (about 70 people in Palo Alto, founded by ex-Adobe and Stability folks) shipped Reve 2.0 on June 3.

Its whole trick is "layout-first". Instead of rolling the dice on one finished image, Reve plans the picture as an editable, almost code-like layout before it renders the pixels.

So you can move, resize, recolour and rewrite individual elements and keep iterating, without re-generating the whole thing and praying.

Reviewers keep calling it "a work file you iterate on, not a one-shot slot-machine pull."

Plus native 4K, strong prompt adherence, and clean text inside the image.

It's #2 on the blind-preference image leaderboard behind OpenAI's GPT Image 2 and just ahead of Google's Nano Banana 2.

The downsides: it trails the leaders on photorealism, faces and identity; it sometimes drops small prompt details, so proofread any text it renders; and there's no vector/SVG. But it's dirt cheap (~$0.0067 an image) and unusually uncensored.

It's also not in Higgsfield (which carries its rivals: GPT Image 2, Nano Banana, Seedream, Recraft). To try Reve you go direct at app.reve.com or through fal.ai.

Kickbacks: Claude Code has ads now, and I'm getting paid

You know that little spinner that ticks over while Claude Code is thinking?

Andrew McCalip just turned it into an ad network. It's called Kickbacks - advertisers bid to show short text ads in that spinner, and the clauders who install it keep 50% of the ad money. "Get paid to wait."

before:

after:

I tested it as a consumer and an advertiser.

As a consumer, setup took about 30 seconds: install the extension, sign in with Google, done (it works in VS Code, Cursor and the terminal). Then it just runs in the background while you work, and a little meter ticks up showing what you've earned.

I had it going for a whole day and made $8.50, but now it seems to have dropped right down to about $1-2/day max.

The bit I didn't expect to like: I've already found a load of cool tools just by clicking the ads I get served. They all feel like interesting little AI startups, so it doubles as a discovery feed for new stuff worth trying.

As an advertiser, I went the other way and ran ads for this newsletter inside it.

Across three campaigns I've put in about $132. The two that have finished brought in 21 subscribers for $77 - about $3.70 each - and a third ($55) is still delivering.

A big one-shot buy mostly shows the SAME developers the ad more often. It doesn't reach proportionally more new people, so the extra money just bought repetition, not reach.

Here’s the link if you want to try it yourself.

Also this week...

• SpaceX just bought Cursor → for $60 billion! xAI folded into SpaceX a while back (in case you were wondering why a rocket company bought a coding agent platform).

• Salesforce is buying Fin (Intercom's AI support agent) → everyone wrote Salesforce off, but I’m not so sure. Once you've plugged all your apps into Claude, you barely open the app front-ends anymore. It's all about how good the backend/API is so an agent can drive it, and that's where Salesforce beats a prettier rival like HubSpot. Between their new CLI, their MCP, and now buying the #1 AI support agent, they're playing the AI age smarter than people give them credit for.

• Graphify → a free, open-source tool that gives Claude Code a cheap knowledge-graph memory of your codebase. It answers repo questions for ~40% of the usual token cost by reading a map instead of grepping every file. Really sick concept, and Chase breaks down the setup perfectly.

• Perplexity gave its agent a memory → it's called Brain. it plugs into Computer (Perplexity's agent that goes off and does tasks for you) and learns from every task you run, so it stops starting from scratch each time and gets sharper the more you use it. research preview for Max subscribers.

• Claude Design got some upgrades → it now stays on-brand with your design system across projects, lets you edit right on the canvas, syncs with Claude Code, and plugs into more of the tools you already use. nice if you design with Claude.

💡 Builder's notes

Every agent harness is the same thing.

If you've followed me for any length of time, you'll know I bang on about this: all the agent harnesses on the market are basically the same thing. Manus, Perplexity Computer, Claude Cowork, Claude Code, Codex, the lot. They're all just really capable general/coding agents you can point at almost any knowledge-work or dev task.

What actually turns one into a real employee comes down to three things:

• Context - information about what you do, who you help, how your business works. without it the agent's useless and you're explaining yourself on a loop.

• Tools - hook it up to your email, Notion, Slack, Stripe. no tools, no tasks actually getting done.

• Skills - SOPs for AI. you train it on your repeatable processes: step one do this, step two do that.

Open up any of these harnesses and it's the same three things every time. A project area to dump context (same as attaching files to a custom GPT or a Claude project). A "tools" or "connectors" tab to one-click log into your apps. And some way to build and manage skills. That's it.

So which one should you use? It really just comes down to two things: how customisable it is, and how hard it is to use. You can plot them all on a graph like this:

For example, I had a play around with Base44's superagents yesterday and it's a really cool tool. SUPER simple interface, but sure enough: a tab for context, a tab for connectors (tools), and a tab for skills. Same three things, every time.

There was no option to pick a model or really any other config, which I don't think is a bad thing at all. there's a huge market for exactly that, people just getting started with agents who want the simplest possible way in. that's why it sits where it does on the chart: about as easy as it gets to start using, but very low on customisability.

The analogy I always come back to: learning to drive. Becoming AI-literate is learning to drive, and these harnesses are just different cars. Once you know how to drive (how agents work, how to give them context, tools and skills) you can hop into any of them and be flying within five minutes.

The indicator might be on the other side, but you'll figure it out fast.

Work amplification vs work automation.

I caught up with a mate for coffee in New York a couple of weeks back, and he gave me a name for something I've been banging on about for ages (the personal AI OS vs business infrastructure thing). His version is way cleaner: work amplification vs work automation.

Work amplification is the AI operating system I’ve built for myself. it's an extension of me as a founder, custom software where agents do the work and I sit in the cockpit all day directing them. 10 to 100x leverage.

For me that's Claude Code and VSCode, and it's all the day-to-day stuff: plan a marketing campaign, bang out an email flow, spin up a landing page, prep for a podcast.

Work automation is the agents running real business processes that don't belong in your personal ecosystem. the obvious one is a customer support agent. you don't want that running out of your own setup with skills you have to babysit.

You want it on something more autonomous, and that's exactly where harnesses like Hermes and OpenClaw come in.

The thing I cracked this week: sharing skills across a team.

Building good skills is the single most important thing to learn right now. you can automate basically any business process with a great Claude skill, or a few chained together. I’ll die on that hill.

The problem: there's no clean way for a team to share and collaborate on skills. save it as a file and airdrop it to your copywriter? the second you both make changes, you've got two diverging copies. Google Drive or Dropbox?

No live syncing, and a mess of its own. an Obsidian vault full of skills? great solo, but honestly I've mostly stopped using Obsidian, it's brilliant for one person and a nightmare the moment a team works in it. they're local files, so collaborative editing just shits itself.

What actually works: make a GitHub repo called "skills" (or one per department, marketing, finance, etc.), then turn that repo into a marketplace plugin. your whole team just runs /plugin to install it and they've got every skill in the repo.

Someone improves a skill, they tell Claude Code to update it, it pushes to the repo, and it auto-updates on everyone's machine. no pulling, no version drift. you also get proper version control, so when a new intern scrubs a skill, you just roll it back.

(hope I didn't lose too many of you beginner folks with that one.)

🧰 Tools to try

• drop.new (by Vercel) → ok this one's just fun. drag a file or folder into your browser, and boom: live production URL, instantly. No need to setup or deploy. If you've ever one-shotted a little page or game with Claude and thought "now what", this is the now what. (also: one engineer built the whole thing. mad.) → Guillermo Rauch's post

• here.now → free, instant web hosting for your AI agent. you tell it to publish something, a page, a doc, a dashboard, a little prototype, whatever it just made for you, and it's live at a real URL in seconds. no account, no repo, no deploy step. anonymous sites are free and live for 24 hours, sign in to keep them forever.

• primitive.dev → email, but for your AI agents. it gives an agent its own email address and a simple way to send and receive mail, with incoming emails arriving as structured data it can act on. so your agents can email people, or kick off work from an email they receive. free tier to start.

• Flow → a tiny Pomodoro timer I've been using lately. nothing fancy, but it's really helped my productivity.

Some mysterious person is still logged into my Youtube account smh…

I promise it’s not me watching 15 hours of Minecraft.

Cheerio chaps,