ISSUE 02 · AUGUST 11, 2026
AI is all we do. We keep up so you don't have to.
WHAT WE'RE SEEING IN THE MARKET
Spending more on AI right now might put you ahead.
We keep hearing the same worry in workshops about AI token leaderboards, whether spending big on AI usage is smart or reckless. Here's what actually surprises people when we answer: a company spending more right now, on the right things, ends up ahead of one holding back to prove ROI first. The number was never the real signal. What matters is whether your people know why they're spending it, and whether anyone's watching where it quietly leaks.
THE WHY
Your bill has three meters, and you've only ever seen one
A token is just a chunk of text, usually smaller than a word. Every AI system reads and writes in them instead of sentences. That's the boring part.
Here's what actually matters: every time your team sends a prompt, three different meters start running.
Input. What you typed, any files you attached, the whole conversation so far. The cheapest meter, per token.
Reasoning. The model thinking to itself before it answers. You almost never see this, and you're still charged for it, at the expensive rate.
Output. The actual answer you read. Priced higher still.
Most people only ever look at the third meter. The other two are why the bill never matches what anyone remembers typing.
It gets worse. In April, Anthropic shipped a new version of Claude at the exact same price per token as the one before it. Same sticker, same $5 per million input tokens.
But the way it counts tokens changed. Anthropic said so directly in their own release notes: the same input can map to 1.0 to 1.35 times more tokens depending on the content. Independent testing by developer Simon Willison found some requests using close to 1.5 times more tokens for identical text.
If you were only watching price per million tokens, you'd have sworn nothing changed. Right up until the invoice did.
I think most mid-market leaders are getting surprised by bills they can't explain, because nobody ever told them reasoning tokens exist, or that one ambitious prompt can quietly cost as much as a hundred simple ones. When the bill lands anyway, the conversation skips straight to cutting usage, because that's the only lever anyone can see.
The real conversation is different. Spend on purpose, and tell your people why, so the money buys real work instead of confusion.
YOUR MOVE
Ask your team the question that actually surfaces the problem
Skip the budget spreadsheet for a minute.
In your next team check-in, or a message in your team channel, ask one question: "When was the last time you didn't use AI on something because you were worried about how it'd look, or what it would cost?"
The number matters less than the pattern. Listen for people mentioning:
Holding back on a task
Second-guessing which model to use
Quietly skipping something because it felt too expensive
That's the real cost of an unclear AI budget: people doing worse work to avoid a bill they can't explain, not employees who are overspending.
Once you hear that, the follow-up is simple. Tell your team plainly that time spent exploring what AI can do is how people get faster and better at using it. That's the whole point of giving them room to experiment.
Then tell them where the real waste actually is. Automations nobody's checked in weeks. Long chats left running after the answer already showed up. The biggest model doing a job a cheaper one would handle fine.
Give people permission to spend on the first kind of thing. Put someone in charge of watching the second.
WHAT TO SKIP
Don't lock every account down with a hard cap yet
Some leader reading this is going to want to solve token anxiety by capping every employee's usage this week. That solves the wrong problem. It just moves the fear from "I might get in trouble for this" to "I'm not allowed to try this at all," and now nobody's experimenting either way.
A hard cap makes sense for the one or two automations you find that are genuinely running wild. It's the wrong first move for the fifty people you're actually trying to get comfortable using AI in the first place.
THE ROUNDUP
Signals from the noise this week.
Claude Token Counter, now with model comparisons (Simon Willison): The measurement behind this issue's central point. Willison found Claude's April tokenizer change pushed some requests to nearly 1.5x more tokens for identical text, at an unchanged price per token. If you want to see the receipts, this is them. Read it here
Introducing our Artifacts Hub and Adoption Dashboard (Nathan Lambert, Interconnects): A free way to check whether a cheaper open model has actually caught up to what you're paying for. Worth five minutes before your next AI vendor renewal, whoever the vendor is. Read it here
Import AI 467 (Jack Clark): Over 1,100 employees across OpenAI, Anthropic, Google DeepMind, and Meta signed the same letter asking Washington to help build tools to deliberately slow AI progress down. Worth knowing regardless of your politics on AI regulation, because it's the industry's own people saying the pace is a live concern. Read it here
Introducing OpenAI Presence (OpenAI): What a real agentic customer support deployment actually looks like once you get past the demo, including the parts OpenAI is charging enterprise services fees to help set up. A useful reality check if anyone's pitched you an agent as something you just turn on. Read it here
Not sure whether your AI spend is buying real work or just quietly leaking? Let's look at it together.
That's issue two. Hit reply and tell me where your team's AI budget conversation is stuck. I read all of them.
Justin
Why of AI