I saved an employee 10 hours a week

whats up Y'all

big theme for this month seems to be new models, every major company is trying to get a slice of AI at the moment

anyway once again here is everything you need to know ;)

---

📌 TL;DR

• ChatGPT just got insanely cheap → the budget model dropped 80% in price, and from today it's free and unlimited for everyone
• Fish Audio clones a voice from five seconds of audio → and the API is free until 31 August
• How to prompt Opus 5 → Anthropic's own four rules, and three of them tell you to stop doing something
• I saved this employee 10 hours a week → bots pull 25 builders' stock lists into one spreadsheet, five times a day
• A free Wispr Flow killer → plus Anthropic building its own chips, and Meta's new coding agent losing on its own charts

---

ChatGPT/Codex is insanely cheap

On 30 July OpenAI dropped the price of its cheap model (GPT-5.6 Luna) by 80% and its middle model (GPT-5.6 Terra) by 20%.

And it gets even better. Starting today (August 7th), GPT-5.6 Luna is free and unlimited for text chats on the Free and Go plans. Limits still apply to file uploads and images, but the everyday chatting is now genuinely unlimited.

GPT-5.6 Luna is comparable to Claude Sonnet 5 intelligence wise, so honestly pretty good.

---

Fish Audio can clone a voice with five seconds of audio

Okay, so Fish Audio just raised $52 million and shipped a new model called S2.1 Pro on 28 July, and it clones a voice from roughly five seconds of audio.

They say it runs twice as fast as Cartesia and costs a sixth of what ElevenLabs charges. It covers 83 languages, and you get word level control, so you can direct how one word lands instead of regenerating the whole clip and hoping.

Here's the bit that actually matters. The API is free until 31 August. No card, no waitlist. You can try the voice clone here without any code.

---

How to prompt Opus 5 (the four rules)

Okay, so Anthropic quietly published a guide on how to prompt Opus 5, and three of the four rules are things you need to stop doing.

That is the part worth sitting with. Most prompting advice tells you to add more. This one mostly tells you to delete.

So here are the four things you need to do explained super simply.

1. Give it the whole job at once.

The old advice was to break a task into steps and feed them one at a time. Anthropic now says the opposite. Their words: it "performs best when given the complete task specification up front and left to run."

So instead of walking it through a task in five messages, write the whole thing in one, then let it go.

2. Say what you do not want.

Opus 5 has a habit of doing more than you asked. This is kind of what I was saying last week. Anthropic's own wording is that it "can also expand the scope of a task, adding steps that weren't requested."

The fix is one line at the end of your prompt saying what is out of scope. You asked for a summary, not a plan. You asked for three options, not five.

3. Tell it how long the answer should be.

Two different lengths, and people only ever set one. There is the length of the reply it types back to you, and there is the length of whatever it actually produces, like a document or a report. Set both, or it will pick for you.

4. Delete "check your work" from your prompts.

This is the one almost everyone is guilty of. That line you stick on the end, telling it to double check everything before it finishes.

Anthropic says taking it out "reduces wasted tokens with no loss in quality." Opus 5 already checks itself. You are paying for it to do the same job twice.

Anthropic's full guide is here if you want to read it yourself.

---

Also this week...

Wispr Flow released "Notetaker" → the dictation app now takes your meeting notes too, with real names instead of "Speaker 1" and summaries built around decisions and next steps. Granola in shambles.

Anthropic is building an in-house chip team for Claude → they're hiring engineers to design custom chips alongside the models themselves, so Claude runs faster and cheaper at scale. Same play Apple ran with its M-series chips.

DeepSeek released V4 Flash → a very fast, very cheap Chinese model comparable to Claude Sonnet 5 or GPT-5.6 Luna. One catch: it's API only, so the free DeepSeek app doesn't have it. Only matters if you pay per API call.

Meta launched its own Claude Code competitor, Muse Code → Claude Opus 5 beat it on every single chart Meta published, including the benchmark Meta built itself. And the model that actually tops the public leaderboard got tested once and left off the rest of the charts. Meta's own AI chief, Alexandr Wang, said outright they're competing on price, not capability.

---

💡 Builder's notes

I saved this employee 10 hours a week

So a little context on my client. He basically sells new-build investment property in Melbourne, Australia. Around 25 builders send him their available stock every week, in every format you can imagine, and none of it lives in one place. He's the guy between the builders and the buyer.

And before he came to me, his listings were landing in twenty different places in six different formats. Emails, PDFs, spreadsheets, portals. He was burning hours every week just manually reading and researching through all of this.

So I made a system where he's got bots that go and pull all of those sources by themselves, five times a day, hand it to Claude, and Claude drops it all into his own personalized spreadsheet.

Pretty cool right, here is a diagram to show you what I mean.

BEFORE & AFTER

Here is the spreadsheet:

This software company just demolished Wispr Flow

If you don't know what voice dictation is, it's basically just a software where you press a button on your keyboard, brain dump and talk to your computer, and it pastes what you said onto the screen. If you're not using this in 2026, I'm sorry, but you're kind of behind 😂

But don't worry. A company called Willow Voice just released an insane, extremely accurate AI dictation for completely free and unlimited (the paid plan is just extra features). They basically killed the need for paid softwares like Wispr Flow.

So if you never made the one I gave out a couple weeks ago, do yourself a favor and just download this one.

It has built-in grammar correction, strips the filler words, punctuates everything, and formats it for whatever app you're talking into.

An email comes out like an email. A Slack message comes out like a Slack message.

Download it at willowvoice.com

(not sponsored lol)

---

🧰 Tools to try

• Design Spells → a library of little UI/UX details that feel like magic, 300+ real examples tagged by type. Its newest entries are hidden details inside Claude Code and GitHub Copilot.
• Ad Copy Skill by Mariah Burner → you give it your product, your audience, and the Reddit communities where your customers actually hang out. It reads hundreds of comments, finds the exact problems, objections and phrases people repeat, then hands you ad angles you can test.

---

💎 Cool finds

• Someone vibe coded a fully interactive 3D human anatomy app → built with GPT-5.6 Sol and Codex, and it started from a single design image.
• Three AIs were asked to paint the Mona Lisa in a browser → Claude Fable 5, Claude Opus 5 and GPT-5.6 Sol all rebuilt Leonardo's actual 500 year old technique on their own.
• How AI works under the hood, explained extremely simply → by Claude.

---

thats it for today folks,

it means a lot that you are reading this right now.

have a great weekend :)

Hugo