ChatGPT’s $1B ad business + Mistral’s €3B raise
+ 7 stories to catch up on and 16 ways to bring your own AI costs down.
Hey friend,
You know what’s missing from a lot of “this AI agent saved me five hours” posts? How much it cost to get the job done. And how much time someone spent checking the result.
An AI task that costs pennies can still become an expensive habit. You’re paying for every attempt, whether it gets the job done or not. That’s why I think understanding AI spend is becoming a basic management skill. A bigger bill could mean your team is getting more done or your agents are getting very expensive at going in circles. You should be able to tell which. Save your spot for Intro to Tokenomics.
Summary
Read time: 6 min
Highlights Of The Week
16 Ways to Cut AI Inference Costs Without Downgrading Model Output
What’s in your AI bill?
Signals
1. ChatGPT Ads hit $1B in annualized revenue run rate in under 200 days, already used by tens of thousands of advertisers
2. ChatGPT-6 Astra scored 99.9% on ARC-AGI 3, its biggest jumps in abstract reasoning, math, and cybersecurity
3. OpenAI says its agents solved the decades-old Navier–Stokes maths problem in 88 hours, with formal verification taking another 17.
4. OpenAI launches ChatGPT Images 2.5 with up to 50% shorter generation times, sketch input and two new API models.
5. Mistral raises €3B in Series D funding at a valuation above €21B, calling it Europe’s biggest private tech fundraising round.
6. Google DeepMind’s AlphaGenome Atlas predicts the molecular impact of all nine billion possible single-letter changes in human DNA.
7. Developers can connect Claude Code to DeepSeek models using its API; the “20,000 developers using it free” claim remains unverified.
What’s in your AI bill?
By popular demand, I’m doing a live session to explain what goes into your AI bill—and where you might be wasting money.
If you’re a founder or work in ops or IT, you’ve probably approved an AI subscription. But once you add tokens, credits and usage limits, working out what you’re actually paying for gets confusing pretty quickly.
We’ll break it down together:
What you’re paying for: tokens, model pricing and usage-based tools, in plain English.
Why the bill keeps growing: what drives costs as your team starts using more AI.
What to look at: how to spot waste and work out whether the spend is worth it.
Bring your questions. No technical background needed. Can’t make it? Register and I’ll send you the recording.
For the people making AI work at work
A private community for ops, IT leaders and founders to compare AI tools, share what’s working and solve the messy bits together.
I’m selecting our first 50 founding members.
Before you switch to a cheaper model, try this.
You might be paying your AI to read the same information repeatedly, write answers nobody needs to be that long, or redo work it’s already done.
Those are all things you can fix.
I’ve put together 16 ways to reduce the cost of running AI, covering what you send to the model, what it sends back, and how often you need to call it.
Each technique explains what to change and what to test, so you can check that the savings aren’t coming at the expense of useful results.
What’s In Their Stack?
This week: Praneeth Pike Founder & CEO of RabbitHoles.
An AI product still needs a lot more than a model to run. Here’s what Praneeth uses to build RabbitHoles and get it into people’s hands. See Praneeth's full stack →
Want visibility on what's in yours? Map it with ELI.work
Stay curious.
Ghita