Prime Ai Solutions
Read time: 4 minutes
Leader, Leader, welcome back.
One AI bill, broken down: 94% went on reading. Just 6% went on writing the answer.
That changes how you should think about AI costs.
This week: what Spotify learned, how to get the most from Fable 5.2, and a quick review that shows if your AI spend is paying off.
LATEST NEWS
Spotify cut its AI coding bill by 90%. You probably can't.
What happened: Spotify uses Claude to write code at huge scale, and reading big files was eating the budget. So a product manager there, Dimitri Mazmanov, built Shunt. It's a free plugin that stops Claude reading any file over 350 lines. A cheaper Google model reads it instead and hands Claude a summary. Spotify reports a 90% average cut in tokens on big files.
Aakash Gupta (Product Growth) tested it on two codebases.
His results are more useful than the headline.
On one, costs fell 87%.
On his own, the plugin cost more.
And inside Claude Code, which already searches for the few lines it needs, the trick added 19% to 36%.
Why it matters: In his test, 94% of the bill was the AI reading.
The output, which looks expensive per token, was only 6%.
My take: Although this sounds like a coding story. It's actually a finance story. Your AI bill is driven by what you make it read, not what it writes.
Pasting a 200-page contract or a full trial balance into every chat is the expensive habit. Point it at the section you need.
His best warning: one early "91% saving" was Claude refusing to do the job. Cheap and wrong isn't a saving. Check the output before you count the win.
Fable 5.2: smarter, faster, still pricey
What happened: Anthropic released Claude Fable 5.2 on 9 November.
It's up to 30% faster than 5.1 and costs $8 per million tokens in, $40 out. That's double Opus 5.5.
Why it matters: On the Remote Labor Index, a test built on real freelance projects, Fable scored 41.3%. Opus 5.5 scored 21.3%.
That's the closest test we have to real paid work!
My take: Use Fable for the hard jobs. A messy reconciliation, a 40-tab model, a contract where one missed clause costs real money.
Don't make it your default for emails and summaries.
That's paying a partner's rate for a trainee's task!
How to prompt it: Anthropic has published a Fable 5.2 prompting guide. Three points matter for non-developers.
It takes your rules literally.
Old prompts full of "never" and "always" make it rigid. Swap "never change the numbers" for "ask me before changing any numbers".
On lower effort settings, it can stop early and list what's left instead of doing it. If that happens, tell it to finish the job.
Be specific when you say yes.
"Yes, update the forecast tab" beats "go".
The easy route: give Claude your old prompt and the guide, and ask it to rewrite the prompt for Fable 5.2. Then run both versions on the same task and keep the winner.
Better tools, bigger bills. Which raises the question nobody likes answering...
STEAL THIS
The 15-Minute AI Spend Review
Stop asking whether AI saves time.
Ask what you'll do with the time.
Because saved time isn't saved money.
I have a rule: a projected return doesn't count until the business actually captures it.
An example is an insurer:
AI cuts claim summaries from 90 minutes to 20.
On paper, that's worth $1.1 million a year.
But if staff work the same hours and costs don't move, nothing is saved!
So the insurer commits to not hiring six planned adjusters.
That's $540,000 a year. After build and running costs, year one nets $60,000.
$1.1 million on the slide. $60,000 in the bank.
That's the gap most AI business cases hide.
Here's a prompt to run the same check.
First, list every AI tool and project in a spreadsheet: cost, owner, and what it was meant to achieve.
You are a sceptical finance director reviewing our AI spend for the quarter.
I've attached a list of every AI tool and project we pay for, with its cost, owner, and the target we set when we approved it.
For each line:
1. Work out the full annual cost, including staff time spent running and checking it.
2. Put the claimed benefit into one bucket: "captured" (shows up in costs or revenue), "capacity" (time freed but not yet turned into savings), or "unproven".
3. Count only "captured" as a return.
4. Recommend one action: fund more, test again, or stop.
Then give me a one-page summary: total spend, total captured return, and the three lines I should challenge first.
Rules: If a line has no target or no evidence, mark it "unproven". Do not estimate benefits I haven't given you. Finish with the questions I should ask each owner.Run it every quarter.
End each review with three decisions:
what gets more money, what needs another test, and what stops.
SIGNAL / NOISE
Signal: Anthropic launched Claude Sheets, Whiteboards and Meetings in beta. Meetings puts Claude in your Google Meet calls as a visible participant. Before Claude joins a client call, check your engagement letter and get the client's OK.
Signal: Launch prices are teaser rates. Google's Gemini 4 Argon is $2 per million tokens in until 30 November, then doubles to $4. Its sibling, Gemini 4 Carbon, launched at the same intro price. Budget on the standard rate, not the launch one.
Noise: The Anthropic IPO. It started marketing to investors this week, with a listing possible as soon as next week. Huge story, zero change to how you should use Claude this month.
Here's the rule I use: hire AI like you staff a team.
The trainee model handles summaries and first drafts.
The manager model handles analysis.
The partner model (Fable) gets the work where one mistake costs money.
Most teams pay partner rates for trainee work.
Spotify's fix was the same idea: send the reading to the cheap model, and keep the expensive one for judgement.
Want to know where your team's AI money goes, and whether it's paying off? That's exactly what I help companies with.
Hit reply and tell me what you're spending on. I'll tell you where I'd look first.
-Umar Prime AI | primeai.solutions
