the most useful skill I’ve ever built

📌 TL;DR

• GPT-6 Astra → OpenAI’s new model landed as I was publishing. The benchmarks are sickening.

• Fable 5.1 → Anthropic’s Opus 5 killer is faster, cheaper and more natural. I’d test it for large business & knowledge work projects.

• Historic day for AI video → You can now generate video faster than you can watch it

• Builder’s notes → The most useful skill I’ve ever built (must see), paper continues to be awesome, and I had codex help with video editing

GPT-6 Astra oh my lordy lord

GPT-6 Astra landed while I was literally about to press publish, so I'm squeezing in what you need to know now.

I'll do a proper breakdown and vibe check next week.

OpenAI says Astra is its new frontier model for computer use, browsing, coding, research, and polished work like documents, spreadsheets, and presentations.

Access is limited to a small group of organizations today. OpenAI says Plus, Pro, Business, Enterprise, API, and AWS access will roll out over the coming days.

Astra and Fable 5.1 cost the same for smaller normal jobs, but on very large jobs Astra starts to get a lot more expensive (due to a higher cached prompt cost)

On a task with 1 million input tokens plus 100,000 output tokens, that works out to about $27.50 with Astra versus $15 with Fable.

These benchmarks make me sick.

full breakdown next week ya heard 👀

Fable 5.1 is an Opus 5 killer

general + coding

It’s better at coding than the original Fable, communicates much more naturally than recent Claude models, works faster and uses fewer tokens.

knowledge work

This bad boy opens the door to big, long-running knowledge work projects: slide decks, docs, spreadsheets, etc.

Opus 4.5 and the GPT-5.3-era models did the same for vibe coding.

Before them, you had to babysit the model through every build. You’d prompt it, check the work, redirect it, then prompt it again.

With those models, you could give the AI a proper plan, kick it off, and come back hours later to something close to finished.

You still need all that back and forth for knowledge work.

Building a slide deck, writing a report, or pulling together a research project usually means working beside the model. You ask for one part, review it, tweak it, then move on.

Fable 5.1 lets knowledge workers work the same way vibe coders already do. Give it a clear task, let it work for hours or days, then come back to a finished result ready to review.

how to use

At its highest effort setting, you give Fable a substantial project, let it put its head down for hours or days, then review the completed build.

It isn’t especially conversational at that setting. Lower effort settings are better for interactive back-and-forth.

In Every’s internal company-work agent, Fable averaged roughly half the tokens and took half the time as Opus 5.

That’s less processing and less waiting, which makes these large delegated jobs faster and cheaper to run.

writing

Opus 5 and Sonnet 5 were so bad at writing that I moved all my writing work to GPT-5.6 Sol, and it’s been incredible.

Fable is finally a Claude model worth experimenting with again.

If Opus 5 is your daily model, I’d switch today. If you’re using GPT-5.6 Sol for writing, I’d start experimenting with Fable again.

benchmarks vs vibes

The benchmark chart looks brilliant, but Opus 5 also looked brilliant on paper and was awful to use.

Models are pretty much rated on vibes now. How they communicate, the judgment they show and whether the finished work is actually useful matter more than another leaderboard win.

Historic moment for AI video generation

I want you to listen to this next sentence carefully.

You can now generate video faster than you can watch it.

Fal took MiniMax H3, retrained it to follow prompts better and produce stronger visuals, then optimized the machinery running it.

It takes roughly 2.5 to 3 seconds of backend processing for a 5-second clip.

Someone has already built a real-life version of interdimensional cable, an infinite TV channel that generates each scene while you're watching.

We've also got an infinite choose-your-own-adventure game and an infinite TikTok-style feed that generates new videos as you scroll.

I can see a world where media becomes a sort of living organism, generating itself as you watch it.

crazy

Also this week...

• OpenClaw 2.0 landed → it rebuilt the setup and browser app so you can get a Claw running faster, then added shared cloud sessions so someone else can join or take over a task without losing the context. I haven't heard amazing reviews so far.

• Claude can now control desktop apps in the background → from Cowork or Claude Code, it can click, type and work inside approved apps while you keep using the computer. 

• Cursor is losing direct access to OpenAI models → OpenAI's proposed cutoff date is November 12. Officially, it says it can't trust SpaceX to follow its terms. Feels like another chapter in the Sam Altman and Elon Musk beef. Very, very annoying because easy access to every major model was one of Cursor's best features.

• Grok Bot routines can now wake up from webhooks → an external event, such as a Slack message or GitHub notification, can start a routine instead of waiting for a schedule.

• Grok Bot added shareable Bots → you can publish a Bot configuration for someone else to install.

• Wispr Flow now mutes other audio while you dictate → this has been an absolute game changer for me this week. I have music playing, and then whenever I use Wispr Flow, the music pauses, and when I stop, the music plays again (you have to turn it on in settings).

💡 Builder's notes

Paper is now in my core AI stack.

Paper.design has become one of those tools I reach for almost every day.

This week I was building the landing page for Skills Atlas. I gave Claude the project, fed it some reference material, and asked it to spin up a few directions in Paper.

It came back with landing-page concepts I was so happy with.

I got it to do the branding too.

Paper is basically a design canvas for your AI agent that works with code. You connect it to Claude, describe what you're building, give it references, and ask for a few variations.

That's basically the whole workflow. Good context, a few directions, then keep talking through what you like until you land on something.

I then cold emailed the Paper team and demanded a partnership. They happily obliged, so Paper is now the first official sponsor of this newsletter. Give them a round of applause.

Go check out Paper here.

Probably the most useful skill I’ve built…

Ask an agent to improve a skill and it often starts piling on new rules: “don't do this, don't do that.”

Do that enough times and the skill gets ginormous. You end up with overlapping instructions and the whole thing gets sloppy.

I hit this while working on my Content Principles and Reel Script Writer skills.

After I'd made a bunch of corrections to a reel, the agent wanted to bolt more rules onto the script-writing skill. The mistake had happened one layer earlier. One line inside Content Principles needed to change.

I got the idea for a better fix while listening to a Founders podcast about Elon Musk's obsession with first-principles thinking.

The result is a skill called /session-skill-audit.

It reconstructs the complete Claude or Codex session, even after the app has compressed older parts of the chat to make room.

Then it traces the failure back to the first instruction that caused it and proposes the smallest change there.

Now the fix lands in the instruction where the problem began, without another pile of rules getting added somewhere downstream.

Codex for video editing

I've been editing the upcoming course myself, and I think that's a necessary part of building the product.

Editing shows me exactly what needs cutting, which explanation doesn't land, and where I need to re-record something.

Sometimes half the lesson can go. Sometimes I taught a part badly and need to slot in a clearer version.

I want to stay close to that.

I did however, automate the repetitive caption work.

Once I've finished the edit, I run a reusable skill. Codex uses HyperFrames to add the captions, fix difficult terms like CLAUDE.md and AGENTS.md, and apply the same style across every video.

It also spots when a caption covers something important on screen and moves the text somewhere safe.

I know CapCut can add captions. I wanted the AI to understand the screen as well as the transcript.

pretty neat.

🧰 Tools to try

• Paper.design → A design canvas your AI agent can read and edit. Connect its MCP to Claude, give it your project context and references, then ask for a few directions. (Paper is this newsletter's first official sponsor 🎉)

• Lieflat Charts → An agent skill that turns your data into polished interactive HTML charts or full reports. Feed it a dataset and ask for the format you need. (free for non-commercial use, commercial use needs permission)

• YC startup-name generator → A tiny model trained on 6,194 YC company names. Use it when naming your next product has cooked your brain. (runs in the browser, no LLM or backend, according to its creator)

🥣 Brain food

• Rhys's Notion control board for humans + agents → 4 columns, one daily Grok Bot nudge, and a very steal-able way to stop agent work disappearing into chat.

• Codex agents working in a tiny office → an independent prototype where agents wander around and drop finished work in a mailbox. completely unnecessary. I want it immediately.

• Heresy by Paul Graham → Paul Graham on the return of opinions you can be punished for expressing. spicy, but worth reading.