Good morning everyone, happy Friday, and an interesting finding from across 181 studies it that most people who agree with you will never show up.

A paper in Science Advances dug into a huge set of protest research to work out who actually shows up to march for the stuff they say they care about. The insight—being angry barely mattered. Feeling wronged or hard done by the government didn’t either. The second and more interesting insight—the people who do show up belonged to some kind of group and they spent time on social media.

AKA, pain doesn't always get people off the couch, but belonging does.

In product we often think that if the pain is real, people will act. But "who has this problem?" isn't the whole question. We should be asking where those people are and who they hang with, and who there is already doing something about it. Anyway…

In the next 5 minutes over your coffee: the price of AI just got cut in half, the Meta agent that beat ChatGPT's launch numbers and then got kicked off Amazon, Waymo's newest 13-year-old customers, why AI food photos make people less hungry, and a boring $12.55B company you’ve never heard of.

On Tuesday we got Opus 5.5. About an hour later, we also got GPT-6 Sol and GPT-6 Luna. That just can’t be by accident, and looking at the pricing, it’s certainly not.

Sol costs $2 per million input tokens and $10 per million output. That's half of what GPT-5.6 Sol cost, and half of Opus 5.5's $4/$20. Luna, the little one, is $0.10/$0.50. And OpenAI told VentureBeat these are permanent prices.

Opus 5.5 still wins on the hard stuff. So you’re not getting the best model at half price. You're getting the second tier at half price. Fine for a lot of things.

Tier 2 intelligence is where most of our products AI features actually run, and that got 50% cheaper overnight.

Back in August I said AI labs racing for market share would subsidize you until their IPO filings forced honest pricing. This is that race happening in real time. Enjoy it.

Split your AI calls into two buckets: volume work and hard work. I’d move the volume to Sol or Luna and keep the harder stuff on the top models. Then re-run your unit economics. If one launch can move your margins 10 points the next one can move them back.

So AI I got cheaper this week, but my time didn't. Neither did yours.

You've heard me talk about Tendem before. It's a bench of 10,000+ vetted human experts you can hand work to without leaving your AI chat. They just put a brilliant cost calculator on their site—it’s fun to play with (and a great marketing idea you can steal btw). You pick your role, enter what your hour is worth, and add the tasks you keep putting off.

I picked Product at $200/hour and added one competitor and market landscape. Doing it myself takes about 6 hours, which is $1,200 of my time. In Tendem it's $45, or $20 on the starter plan, plus about 19 minutes of me reviewing it.

There's no sign-up to try it. Add the tasks you've been dodging and see if handing them off is worth it for you too.

Meta launched Muse on Sep 8th. At first people laughed and I can’t say I didn’t have my doubts. But don’t ever underestimate Meta’s distribution. Muse is growing faster than ChatGPT did at launch, and it isn't close. In its first 12 days it got 1.8M iOS downloads in the US and Canada, against ChatGPT's 1.3M over the same stretch. It hit 642K US daily users to ChatGPT's 231K, and it's now #1 on the App Store.

On day 12, Amazon blocked it. Amazon says the agent doesn't identify itself when it browses and it seems to store customer logins. Amazon has already sued Perplexity and blocked shopping agents from Google and OpenAI.

Same week though, and Expedia did the opposite. They want to let Muse book trips directly. And on Sep 17th, Mastercard started giving AI agents their own virtual cards with spending caps so bots can buy things without asking

Amazon makes its money from ads and from you landing on its pages, and an agent skips both. Not good for that business. When Walmart tried OpenAI's Instant Checkout, it converted at a third of the rate of Walmart's own site, with smaller carts, so Walmart bailed. Expedia makes its money on the booking fee, not your eyeballs, so it's happy to take demand from anywhere.

If you earn on attention, agents are a threat. If you earn on the transaction, they're just new customers. Pick an agent policy based on the answer to which you are before a platform picks it for you: welcome agents, make them identify themselves, or block them. Ben Thompson’s full "aggregator v aggregator" piece is a good take on who wins agent shopping 👇

A few of my recent Fridays opened lower than usual. So I connected Claude to the beehiiv MCP and asked it to compare my best editions against the weaker ones. It pulled open rates, preview text and audience settings for 16 Fridays in a couple of minutes.

What it found:

So I asked Claude to set up an A/B test on today's subject line, and it wired it up in beehiiv for me. Then I asked it to write this ad from everything we'd just found and put it into my draft. It did that too.

And I did all of it on my phone, waiting for my plane to take off.

That's what you get from a creator platform with proper analytics that your AI can actually work with: answers instead of guesses, and the busywork done for you. Since moving to beehiiv, my open rate has climbed from 36% to over 50%.

Use THEDIFF30 for 30% off your first 3 months.

Teens (13 to 17/yo) in Nashville can now call Waymo’s on their own as long as the account is linked to a parent's. Nashville is the second city after Phoenix, and Waymo says it’s the "first of many."

The person riding and the person paying are different people, and Waymo is selling to both. Teens want freedom. Parent wants convenience. And for a parent, I guess a robot with live trip sharing might be an easier yes than a stranger in a Corolla.

It's a habit play too. Remember the study from last month: we don't really choose things, we choose them again. A kid who gets to school in a Waymo at 14 will be very hard to switch at 25 when self driving cars are the norm.

Classic buyer vs user. For you, who pays for your users in your market, and what does that person need before saying yes? Usually it's visibility and control like sharing, receipts and limits. Build that yours account have two happy customers instead of one.

There’s a slop frenzy of restaurants/delis etc putting AI photos on their menus. It's going really badly…look at this 😂

Back in 2024, Oxford professor Charles Spence found people preferred AI food images to real ones, until they were told the images were AI. After that the exact same images were rated less appealing. He also noticed that when AI "improves" food, it loves adding fat cues, like an extra slab of butter on the mash.

AI images can work right up until people find out it’s AI…and people always find out. "Almost real" is also the worst place to land. Real photos work and so do obvious drawings. The uncanny middle ground loses.

This goes past menus. Anything people eat, wear, live in or hand money to sits in the same zone. It's what I said two weeks ago about AI decks—polish used to prove effort and now it can count against you.

Get rid of AI images on your side. Swap them for real photos, even phone ones. If you want to keep AI art, make it obviously stylized so nobody feels tricked.

For my app Little Moments, I was using stock video from Pexels for the app video montage. Julia told me to use real videos from my phone, I did, and it’s so much better even though worse quality. Every image I use on my landing page, I took.

Temporal makes software that lets long-running AI tasks finish even if something breaks halfway through. Nobody's making TikToks about it though and it just raised $550M at a $12.55B valuation, up 2.5x from $5B seven months ago.

An agent running a code migration or processing an insurance claim can run for hours across dozens of tools, and any one of those tools can time out. Something has to remember where the task was and pick it back up. The more work AI does, the more that something gets paid.

Picks-and-shovels. When everyone's building agents, sell the thing that keeps agents from breaking. Nobody gets excited about it which is part of why it's such a neat business.

If you're building with agents, what are the ways your runs fail? Retries, lost state, times outs, missing human approvals? Each one is a feature. If you're hunting for a startup idea, each one might be a company.

🛠️ A toy—Stacks, from Shibuyaaa, a new software company started by Spotify's head of innovation. It's a game you play with music. Just go have fun with it → Play it

📺 A video—Not sure how the hell I ended up here, but I just watched a 16-minute video on how long I'd last in every prehistoric era. Odd, but I couldn't turn it off → Watch it

📬 A read—A short one I keep thinking about: become the person you'd like to be around → Read it