What's up gang,

Hope you've all had an amazing week. I'm late writing the newsletter again - it’s accidentally become a Friday newsletter lol.

48/60 spots gone so far for the AI course. Early bird offer ends today!

Access is only for the waitlist right now, but I will open it up to the rest of you lovely people + social platforms tommorrow - spots will fill fast.

📌 TL;DR

  • Sonnet 5 → The new mid-tier is a big leap, near-Opus quality at less than half the price. use it for the low/medium-effort grunt work, and keep Opus for the properly hard stuff.

  • OpenAI's GPT-5.6 → A new three-model family (Sol, Terra, Luna). Terra matches GPT-5.5 at about half the cost. unfortunately it's locked to ~20 partners, so you can't use it yet.

  • Fable 5 is back → The ban's lifted, but it's only free on your plan until July 7. if you've got big builds waiting, now's the time.

  • Builder's notes → The subscription-audit skill that tracks every dollar I'm paying (turns out it's $1.6K/mo), plus the two skills that now run my email marketing end to end.

Sonnet 5: near-Opus quality at less than half the price

Claude Sonnet 5 landed on June 30, the first proper Sonnet upgrade since 4.6, and it's a big leap.

It's now landing right near Opus on a lot of real work. only about 2% behind Opus 4.8 on computer use and agentic coding (agentic just means the AI going off and doing multi-step tasks on its own - instead of just answering a question), and actually ahead on knowledge work.

The one real gap is a coding benchmark called SWE-bench Pro, where it sits at 63% to Opus's 69%. that's just a test of how many real coding tasks each one solves solo, so on your toughest builds Opus still wins.

Alongside it, Anthropic put out an official chart plotting cost against performance across the different effort levels, and it's turned "which model should I use?" into a real case-by-case call.

For some tasks Opus 4.8 is both cheaper AND better. for others Sonnet 5 is the smarter pick.

These two charts show it at a glance: Sonnet 5 is the better buy down at low/medium effort, and Opus 4.8 pulls ahead once you're paying for high effort.

Anthropic's official cost-vs-performance-by-effort chart, agentic search (BrowseComp)

IMAGE: Anthropic's official cost-vs-performance-by-effort chart, computer use (OSWorld)

Where Sonnet 5 really shines is the routine, scheduled, low-effort stuff.

a daily brief, clearing an email inbox, the jobs that still need doing properly but don't need an ultra-capable model like Opus behind them.

Anthropic's lineup now has four tiers:

  • Haiku - the cheapest and fastest, but the least capable. fine for simple, high-volume jobs, but not much grunt behind it. I pretty much never reach for it.

  • Sonnet - the mid-tier workhorse. lower cost, seriously strong. Sonnet 5 now gets a 1M-token context window (it can hold a huge amount of text in its head at once, think a whole codebase or a stack of docs) and lands close to Opus quality for less than half the price.

  • Opus - the heavy hitter. Anthropic's best reasoning model, built for the properly hard, complex tasks. often the cheaper choice on tough problems, because it solves them in fewer steps instead of grinding away.

  • Fable - the newest and most capable of the lot, a tier above Opus (their "Mythos-class" model, basically their most powerful internal model made safe enough to release). top of the range, and priced like it.

IMPORTANT PLS READ: The one rule worth remembering: Sonnet 5 is a bargain at low and medium effort. but push it to high, xhigh or max and its cost climbs right up onto Opus 4.8's, and at that point Opus matches or beats it.

OpenAI's GPT-5.6: cheaper and faster, but you can't have it yet

OpenAI previewed a new three-model family on June 26, called GPT-5.6. three flavours: Sol (the flagship), Terra (the balanced middle), and Luna (the cheap, fast one).

Terra gives you roughly GPT-5.5-level performance at about half the cost.

GPT-5.5 is already excellent at coding, following instructions and business tasks, so "same quality, half the price" is the bit that actually lands.

Rough pricing per million tokens: Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6.

But you and I can't use them yet. at the US government's request, access is limited to about 20 approved partners, handed out customer by customer, with wider availability promised "in the coming weeks."

Why so locked down? frontier models are now capable enough to carry real cyber-security risk, and (tied up with the whole Fable 5 saga) that's made these releases political.

It's a shame they're this restricted, I get why though. hopefully we all get access in due time.

Fable 5 is back, and the clock's ticking on the free window

Quick one, since I covered the ban itself a couple of weeks ago.

Fable 5 is usable again. the export ban got lifted and it returned globally on July 1, after 19 days offline.

But read the small print. it's only included for up to 50% of your weekly usage, and only on your plan until July 7. that's basically a week.

After July 7 it goes API-cost only, roughly double the price of Opus 4.8, so it stops being "free" on your subscription.

So if you've got any big builds you've been sitting on, now's the time. crank through as much as you can with Fable 5 while it's still on your plan.

(In case you missed it, here's where I covered fable 5.)

Also this week...

  • Google shipped two creative primitives → Gemini Omni Flash (public preview) does chat-to-edit video: generate a clip, then refine it just by talking to it. and Nano Banana 2 Lite is a ~$0.03, ~4-second image model, near-free for fast ideation.

  • Cursor launched an iPhone and iPad app kick off cloud coding agents from your phone, review the diffs and screenshots, and merge, so work keeps moving while you're away from the laptop.

  • Midjourney revealed a hardware pivot V8.2 is finalised and previewable now with the --preview flag, and they've announced their first hardware project: a full-body scanner, rolling out through a physical "Midjourney Spa" in San Francisco. wild.

  • Notion is retiring its Mail inbox app shutdown Sept 22. The reason? “Today, more than half of Notion Mail users manage emails without ever opening their inbox. So, we’re going all in on using agents to run your inbox.”

Builder's notes

The subscription-audit skill: full transparency on what I am actually paying for.

I built a skill called subscription audit.

it goes through my email inbox and my bank statements, pulls the receipts for every new subscription, price change and cancellation, and builds it all out into a Notion database, so I've got complete transparency on every subscription I'm paying for.

I've set it to run on a schedule every Friday, so it just keeps itself current without me touching it.

My subscriptions at a glance: about $1.6K per month total, auto-built by the skill

The full Notion database the skill maintains: every active subscription, its cost and its renewal date

Two skills that now run my email marketing end to end.

All my time and attention this week went on the course launch, and two skills carried the entire email side of it.

First I built an AI with Remy email copywriting skill that captures my actual voice, how I write marketing emails.

then an email formatting skill that connects to Resend (where I send my marketing emails from).

Here's the loop for every email I sent this week: I run the copywriting skill and voice-note my thoughts, tweak the draft myself (rewrite a few bits, add an intro), then run the formatting skill and it drops the email straight into Resend, styled in my house look.

I do one final once-over, hit send, and it automatically logs the email in my Notion Email Calendar with the content and marks it as sent.

My Notion Email Calendar: the whole course-launch sequence, auto-logged and marked sent by the formatting skill

Two skills, and the whole email side basically runs itself now.

🧰 Tools to try

  • OpenRouter MCP server → Lets your agent pick, price and test the right model for a task from live data, instead of guessing off stale training. plug it into Claude Code, Codex CLI, OpenCode or Cursor CLI and let it choose the model itself. (MCP is just the standard that lets an agent plug into an outside tool like this.)

🥣 Brain food

Just rocked up to my Airbnb in Majorca that Claude got us a €2,800 discount on… we got catfished a little bit.

They've definitely used Google nano banana a little bit on those images One of the bedrooms looked huge in the photos, but it's literally the size of Harry Potter's closet.

alas, we have wifi. and thats all a man an ask for.

love,

Keep Reading