I sat with a client on Tuesday who had done everything right.
Good use cases. Real adoption. Skills built and running. Her team was past the demo and into the daily work, which is the hard part, and honestly it’s the part most companies never get to.
And she was getting shut off. Usage warnings every week, usually by Wednesday.
So we opened her settings, and there it was. Her instance had defaulted to Fable 5, which is the most capable and most expensive model in the lineup, and every single thing her team touched was running through it. Summarizing a voicemail. Reformatting a spreadsheet. Tagging expenses.
She was using a rocket ship to go pick up the pizza.
Here’s the part I want you to hear. Nobody made a bad decision. She didn’t choose wrong. She never chose at all. The default chose for her, and the default is expensive.
On well defined, repetitive work, the flagship model is not better. It’s slower, it costs more, and it hands you the same output. The gap between tiers only shows up when the thinking is genuinely hard or the stakes are genuinely high.
Here’s what that looks like in dollars. These are estimates I built from Anthropic’s published API rates as of August 2026. They are not per task figures Anthropic publishes, and your real numbers will move around based on your document sizes and how many revision rounds you run. But the shape holds.
| Task | Fast tier (Haiku 4.5) | Balanced (Sonnet 5) | Frontier (Opus 5) | Fable 5 |
|---|---|---|---|---|
| 500 receipts reconciled against expenses | $1.25 | $2.50 | $6.25 | $12.50 |
| A quarter of sales calls summarized (60 calls) | $0.72 | $1.44 | $3.60 | $7.20 |
| 100 product descriptions written | $0.32 | $0.64 | $1.60 | $3.20 |
| 2,000 CRM records cleaned and deduped | $2.70 | $5.40 | $13.50 | $27.00 |
| Employee handbook drafted from your policies | $0.45 | $0.90 | $2.25 | $4.50 |
| Brand standards, GTM strategy, and a 3 year pro forma | $2.45 | $4.90 | $12.25 | $24.50 |
Look down the columns. The spread is the same every time. Fable runs about 10x the fast tier and roughly 2x the frontier tier, no matter what the task is.
Which is good news, actually. It means the discipline doesn’t have to be complicated. It just has to exist.
Right. And that’s exactly why nobody ever looks at it.
Every number in that table looks like a rounding error, so it never gets examined. Not until the invoice shows up or the shutoff notices start, and by then it’s been running that way for months.
So run it forward. Forty people, reconciling receipts once a week, for a year.
Same work. Same output quality. Same people. That’s a $23,400 swing on a setting nobody ever touched.
That’s the whole thing right there. Volume is what turns pennies into a line item.
The second thing everybody asks me is whether upgrading seats fixes this. I went into it assuming there was a volume discount sitting there waiting. There isn’t, and I’d rather tell you that than sell you the upgrade.
On Claude’s Team plan you’ve got two seat types. A Standard seat gives you roughly 1.25x the usage of an individual Pro account. A Premium seat gives you roughly 6.25x.
Now do the division. At list pricing, both come out to $20 per unit of capacity on monthly billing, or $16 on annual. Identical, either way you pay. Five Standard seats equal one Premium seat in capacity, and they cost exactly the same. There is no bulk discount.
So what are you actually buying with Premium? Two things, and both are worth paying for.
The first is concentration. Usage limits are tracked per member, so five Standard seats can’t lend their headroom to the one person running the heavy job. That capacity just sits there stranded.
The second is continuity. Premium carries a higher weekly ceiling, so your heavy users stop getting cut off in the middle of something.
So the rule is pretty simple. Upgrade your heavy users specifically. Don’t upgrade everybody. Buying Premium across the board means you’re paying a concentration premium for people who were never going to hit the wall anyway.
There’s a belief floating around that you get more capacity if you work off hours. I wasn’t sure about this one myself, so I went and checked. Here’s the honest version.
It was true, for a while. Back in the spring, Anthropic reduced the five hour limits during weekday peak hours, roughly 5 to 11 a.m. Pacific, and they ran a promotion that doubled limits outside that window and on weekends. By May they’d doubled the five hour limits again and removed the peak hour reduction for Claude Code on Pro and Max plans. The current help documentation doesn’t mention peak hours at all.
So it’s documented history, not something you can bank on today.
I’d still schedule your recurring jobs and automated tasks overnight. Not because there’s a discount you can point to, but because shared capacity is demand sensitive and it costs you nothing to be smart about it. Just don’t build your budget around it.
Now the one that got me, because it’s the same disease and I had it worse than my client did.
Anthropic changed how Claude’s memory works this week. It used to be two separate brains — what you taught it in a chat stayed in the chat, what it learned in Cowork stayed in Cowork. Now it’s one system, and you can open it up and read the actual list of things Claude believes about your business. Edit it. Delete what’s wrong.
That last part is what matters. If you’ve ever burned the first four minutes of a session re-explaining what your company does, that’s the tax this removes. And a confident wrong memory — a client you no longer serve, a price you no longer charge — quietly poisons every answer downstream until somebody goes and looks.
So I went to turn it on. And I couldn’t find it.
I looked through my settings twice. No Memory. I assumed it hadn’t rolled out to my plan yet, or that some admin switch was holding it back, and I moved on with my day.
Both assumptions were wrong. Here’s what was actually true.
There is no Memory item in the settings navigation. Not on my plan, not anywhere. Memory is a section — the first one — inside a page called Capabilities. If you’re scanning the sidebar for the word “memory,” you will never find it, because the word isn’t in the sidebar.
Here’s the real path: click your name in the bottom-left corner of the sidebar → Settings → Capabilities. Memory is the first section on that page. On a Mac, ⇧⌘, opens Settings directly, and it drops you on General, which is not where you want to be.
And when I finally got there, the second thing hit harder than the first.
It was already on. Both toggles. “Search and reference chats,” on. “Generate memory from chat history,” on. And a row reading “View and manage memory · Updated 2 hours ago.”
Updated two hours ago. It had been running the whole time, building a picture of my business, and I had never once seen the page.
Then I checked the admin side, where there’s a switch labeled “Enable memory for your team.” Also on — and marked Default. Not “Set by admin.” Default. The interface was telling me, in a single word, that I had never made a decision about this. The platform made it, months ago, and I inherited it.
That’s the whole article in one label.
So when you go look, click the row that says “View and manage memory.” Read the list like you’d read a new hire’s notes after their first month. You’ll find something useful you forgot you taught it, and you’ll find at least one thing that’s flatly wrong. Delete that one. There’s a “Manage edits” button in there too, so you can see what’s been changed and clear it if you want a fresh start.
Ten minutes. Probably the highest-value ten minutes in this article, because unlike the model setting, this one has been quietly shaping your answers rather than just your invoice.
You can do this in twenty minutes.
One. Check your default. Open your model selector. If it says Fable 5 or Opus 5 and your team is doing routine work, you just found your leak. This single change is usually worth more than everything else on this list combined.
Two. Right size your seats. Figure out the two or three people who actually hit limits. Upgrade those people. Leave everybody else on Standard.
Three. Match the model to the task. Fast tier for work that’s well defined, repetitive, and cheap to catch if it’s wrong. Balanced for most real business work, and that should be your default, because it’s the right answer more often than people expect. Frontier only when a mistake is expensive or the thinking is genuinely hard.
Four. Read the pages, not the menu. Memory is the example I just walked you through, but the lesson generalizes: the features that cost you money and shape your output are grouped onto pages, and the page names don’t match what you’d search for. Once a quarter, open every settings page and read it top to bottom instead of hunting for a keyword. You cannot audit a control you can’t find, and “I couldn’t find it” is indistinguishable from “it isn’t there” right up until it costs you something.
For step three, we built you a tool. The AI Model Picker is free, no signup, and it works whether your team is on Claude, ChatGPT, Copilot, Gemini, or Grok. Describe the task the way you’d say it to a colleague, and it names the exact model to open, the reasoning behind it, and a prompt written for your task.
Send it to whoever on your team keeps asking which one to use. It answers the question once.
My client wasn’t wasting money because she was careless. She was wasting money because she was busy, the tool worked, and nothing ever forced her to go look.
I wasn’t missing the memory settings because I’m not paying attention. I run an AI consultancy. I missed them because they were one click off the path I expected, under a heading I wasn’t searching for — and because “I can’t find it” felt like a complete answer at the time.
That’s most of what I see out there. The waste in AI adoption almost never comes from bad strategy. It comes from good strategy running on defaults nobody set, on pages nobody opened.
Go check your settings. It’s twenty minutes, and I’d argue it’s the highest return twenty minutes on your calendar this week.
Keep Climbing!
John
One email every two weeks — practical AI for small business owners. Free, and the forwarding kind of good.
Subscribe free →A note on the numbers. Cost figures here are estimates I derived from Anthropic’s published API pricing as of August 2026 and typical task volumes. They are not Anthropic published per task rates. Seat multipliers and prices reflect Claude Team plan documentation as of August 2026. Settings paths were verified by hand on a Team plan on August 27, 2026 — interfaces move, so if yours looks different, trust your screen over my instructions. Nothing here is legal, tax, or financial advice.