Good evening folks,

The AI course: six live sessions over two weeks, sold out on Tuesday 🍾

I havent been so excited for a project in a while!

📌 TL;DR

  • GPT-5.6 → OpenAI's new family (Sol, Terra, Luna) is out, and Every ran it for a month before anyone else. Their verdict: a Porsche to Fable's warp drive, a much better writer, and the first model that can run whole loops of work on its own.

  • Claude Cowork runs with your laptop shut → hand it a task at your desk, pick up the finished work on your phone. Scheduled jobs now run with no device online at all. Beta, Max plan first.

  • Notion's Agent Hub → run your custom Notion agents from your phone. Native to your workspace and multiplayer, so the whole team shares one agent. Firmly the automation side, not the cockpit.

  • Grok 4.5 → cheap, fast, "a weirdly good default code model" at $2 / $6. Free for a limited time in Cursor and Grok Build. Proof that models matter less and less now.

  • Builder's notes → I built Ledger, my own finance app, and folded every subscription into it. Plus 3 prompts to run before Fable 5 leaves your plan on July 12.

GPT-5.6: the model Every has been secretly running for a month

OpenAI shipped GPT-5.6 today, but the story worth reading isn't the launch post. It's that Dan Shipper and his team at Every had it for about a month before the rest of us. They were one of ~20 early partners running it, and their verdict just dropped.

So I'll let them carry this one.

Dan's headline: "GPT-5.6 is like a Porsche, Fable is like a warp drive." (Fable being Anthropic's monster model, the one that leaves your plan on July 12.) The Porsche is what you drive every day, the best mix of power, speed and performance for normal knowledge work and coding.

Fable is the beast you save for getting across the galaxy. So GPT-5.6 is not a Fable killer, more a division of labour. Run the two together and, in his words, it's "goated."

A few things stood out to him:

  • It's a much better writer than Fable. It one-shots marketing emails that every model before it fumbled, where Fable gets verbose and drifts off into its own private language. And it's a lot cheaper than Fable too.

  • His biggest claim: it's "the first model that can reliably run whole loops of knowledge work, not just individual tasks." He's got it running autonomous loops for his email, his hiring, and his meeting and Slack digests. (A loop meaning a job it works end to end on its own, not a single ask you sit and babysit.)

  • The one caveat he gives: it doesn't reliably catch the coding issues Fable catches. So it's still more of a daily driver than a safety net.

It comes as a family of three:

  • Sol - the flagship. Frontier reasoning, plus a new "max reasoning effort" mode and an "ultra mode" that spins up sub-agents in parallel (several helper models working at once instead of one plodding through step by step).

  • Terra - the everyday one. GPT-5.5 performance at 2x lower cost.

  • Luna - the fastest and cheapest, built for high volume. Should comfortably beat Haiku.

Pricing per million tokens (in / out): Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6.

One important thing, and I don't want to overclaim here. OpenAI says it goes public today and is rolling out, but they haven't confirmed every ChatGPT plan gets it on day one (the preview was API and Codex only).

Expect a phased rollout: API and the higher tiers first, then Pro and Team, then Plus, then Free over the coming weeks. So check whether it's actually live on your plan rather than assuming. When it reaches you, Terra's the pragmatic everyday pick.

Claude Cowork now runs with your laptop shut

Cowork is Anthropic's agent that goes off and does work for you, their answer to the whole harness category.

And here’s a quick teaser if you want to see it in action:

The update: hand it a task at your desk, then pick up the finished output from your phone. Close the laptop and it keeps working, because the sessions now run in the cloud instead of on your machine.

The bit that matters is that scheduled tasks now run with no device online at all.

Set client prep for 6am and Claude works through the threads and transcripts overnight, builds you the brief, and leaves the follow-up drafted but unsent, waiting for your nod. When something actually needs your judgment, the question lands on your phone.

Chat and Cowork are also moving into one home on web and desktop, so your projects and artifacts live in one place across both.

Just take into consideration, it's a beta, rolling out over the next several weeks. Max plan first, with other plans to follow.

(They've also doubled Cowork usage limits through August 5 to mark the launch.)

Notion agents on your phone

You've been able to build custom agents in Notion for a little while now.

It's the same idea as a Claude Code agent: a folder you hand an instruction file, telling it what it's responsible for, how to respond, and which tools and skills it's allowed to use. Just with a nice view inside Notion sitting on top.

What's new is the Agent Hub, an iOS app that lets you run and assign those agents from your phone.

If you're a heavy Notion user and your whole business already lives in there, it's worth a play.

The real edge over doing this in code comes down to two things. Notion agents are wired natively into your workspace, and they're multiplayer: build a marketing agent once and your whole team can use it. That's a lot harder to pull off with a code setup.

Here's where I file it. It's the work amplification vs work automation thing again.

Amplification is your personal cockpit, the setup you sit in all day directing agents through your own work (for me that's Claude Code and VS Code).

Automation is agents running real business processes that don't belong in your personal ecosystem, a customer support agent say, on something more hands-off.

Notion agents fit that second bucket, which makes them a genuine competitor to OpenClaw and Hermes rather than to Claude Code.

Grok 4.5 is really good, but that barely matters anymore

Another week, another model. And honestly, models matter less and less now, because they're all so good.

It reminds me of iPhones. The jump from the 3 to the 4 to the 5 was massive. The jump from the 15 to the 16 to the 17 is marginal. That's about where we are with frontier models. So I won't bury you in benchmarks.

I'd rather tell you where Grok 4.5 actually earns its place.

xAI (now folded into SpaceX, so technically SpaceXAI) shipped Grok 4.5, their first model trained specifically for coding and agents, built together with Cursor. It's pitched as a cost play, not a raw-intelligence flex. Elon's own comparison is roughly

Opus 4.7-level intelligence, but cheaper. $2 / $6 per million tokens.

And right now it's free for a limited time. You can pick it in Cursor (all plans), or in Grok Build, their Claude Code equivalent that runs plugins and skills.

The verdict worth reading is Theo Browne's (Elon reposted it, which tells you something). He calls it "a weirdly good default code model" and "a good alternative to Opus 4.8": the cheap daily driver you run everything through, only bringing a bigger model in to clean up.

Grok 4.5 against the field (Grok highlighted) - but the story here isn't the benchmarks

A couple honest caveats: it doesn't catch everything Fable catches, and it can't orchestrate sub-agents like the newest models can.

Ay if you can fit it in your stack then give it a whirl, but I’m not too fussed.

Also this week...

  • GPT-Live → OpenAI's new voice models (GPT-Live-1 and its mini) listen and speak at the same time. They handle interruptions, throw in the little back-channel cues ("mhmm," "got it"), and stay quiet while you're thinking. They've replaced Advanced Voice Mode as the engine behind ChatGPT Voice, and the mini is the default on the free tier. The reasoning still comes from GPT-5.5 behind the scenes, so it's a jump in conversation quality, not a smarter brain.

  • Fable 5's free window got pushed to July 12 → Anthropic extended Fable 5's included access on paid plans from July 7 to July 12 (11:59pm PT). After that it moves to pay-per-use credits at $10 / $50 per million tokens. If you've been sitting on big builds, you've got a few extra days to run them on your plan.

💡 Builder’s notes

I built Ledger, my own finance app, and folded my subscriptions into it.

This was a proper Fable 5-worthy job I'd been meaning to do for ages, and it nailed it. Ledger is my own finance app, revenue, expenses, the lot.

What i love most is that it now holds every one of my subscriptions.

Every subscription in one place - the Friday audit skill keeps it current, with one-click cancel links

The Ledger overview dashboard (figures blurred) - revenue, expenses, Wise balance and more in one place

My subscription-audit skill runs every Friday on a claude code schedule.

It reads my bank account and my email receipts, works out what's active, what's up for cancellation and what's new, and now it writes straight into Ledger.

Remember that subscriptions Notion database I showed you last week? I folded the whole thing into this app.

There's a subscriptions section with every detail I need, plus one-click links that drop me straight onto each service's billing page to cancel.

3 prompts to run before you lose Fable 5.

Promise this isn’t ‘click-bait’. Fable leaves your plan on July 12. Before you lose it, run these three prompts and squeeze the last bit of juice out of it while it's still included. I put them in a quick guide:

  1. Find Fable-worthy work - a lighter model scans your real projects and backlog and hands you ready-to-run Fable briefs for the top three tasks actually worth the spend.

  2. Delegate a task overnight - hand Fable a job with /loop and it keeps your Mac awake, works away unsupervised until it's done, and you wake up to finished work plus a short list of the decisions only you can make.

  3. Audit your setup and skills - Fable reads every instruction file and skill in your Claude Code setup, hunts down the contradictions, stale rules and dead pointers, and rebuilds it smaller and sharper. It reports first, so nothing changes until you approve.

X bookmarks, synced into Notion.

Someone on X posted a screenshot of a setup that syncs their Twitter bookmarks into Notion. I literally handed Claude the tweet and said "build this for me." Now I've got a Notion database that syncs all my X bookmarks automatically.

It's kind of productivity porn. I don't actually know what I'll use it for yet, but it seems cool.

My X bookmarks, auto-synced into Notion - built by handing Claude a single screenshot

🧰 Tools to try

  • mattpocock/skills v1.1 → Matt Pocock's skill pack got an update: /wayfinder to plan out more ambitious work, plus /to-spec and /to-tickets. (Same author as the /teach skill I featured a few weeks back.)

  • Ditto → clones any website into clean, componentized code in under 5 minutes. Free and open source, with a REST API and MCP server included.

  • The "Taste" skill for Fable 5kills the generic AI-slop and gives Fable the tools to ship properly good-looking design.

  • Goose Ads → a full ad-creative team inside Claude Code. It finds the ads your competitors are paying to run and remakes them for your brand.

  • Maker Skills for AI Agents → 18 free skills for founders and operators (decisions, research, a CFO), plus meta-skills that build more skills.

  • Photo-to-gradient websiteturns any photo into a clean gradient. Handy little design utility to bookmark.

🥣 Brain food

thats all for today folks,

Keep Reading