It’s not a secret that the Labs make more money when you spend tokens. But your company wins when you ship more product and get paying customers.
A casino analogy isn’t perfect. Casinos are built around negative-sum games; AI agents can create real value: a diff, a bug fix, a feature, an afternoon saved. But the incentive rhymes. The Labs don’t need you to lose. They just need you to keep playing: more context, more thinking, more retries, more agents running in parallel. Your ideal tool solves the problem and gets out of the way. A usage-metered tool is rewarded for keeping the session alive.
That’s the token casino: useful software wrapped in mechanics that make spend feel like progress. It starts with the oldest trick in the book: abstract the money.
Technique 1: Chips, not dollars
You don't spend dollars. You spend tokens.
Tokens are like poker chips. Put one layer between you and the price, and it stops feeling like real money. A $100 poker chip feels the same as a $1 one. Similarly, the whole point for the Labs is that you don't think about the dollars when you’re triggering a /goal run.
Technique 2: The UI is a slot machine
I’ve been obsessed with the psychology of addictive product design since I first started coding. Social media apps like Facebook and Snap have done this well for a decade, but next time watch the TUI of Codex and Claude Code. The token counter spins up. The little "thinking…" animates. It feels good to watch the agent work.
There’s this whole style now: retro terminals, pulsing status bars, spinner verbs firing, live token and runtime counters, all praised for keeping you "in flow." Similar to a casino, there's no clock on the wall. The Labs don’t need to make it easy to tell how much money you’ve spent. Internal Microsoft docs for its new "Scout" assistant reportedly listed the first goal in plain text: make people addicted.
Technique 3: The money spent is the flex
If you pay attention on X, they're already turning spend into a score and something to brag about.
Codex "Profiles" show a token leaderboard: lifetime tokens, peak tokens, streaks. It rewards burning tokens, not shipping work. During OpenAI DevDay last year, OpenAI gave out awards for trillion token spenders like Cognition.
These awards are fun and cute, but your finance team is realizing these receipts add up. Uber reportedly spent its entire 2026 AI budget in four months. Microsoft reportedly pulled Claude Code from ~100k engineers.
Getting tilted
A former Codex engineer Calvin French-Owen said every run feels a little like gambling and as someone that has spent too many nights at poker tables and at my desk, I couldn’t agree more. Here's how it usually goes. The agent misses. I try again. It gets worse. I try again. Worse. I give up, get tilted, open a fresh session, and change the prompt: a bigger model, more thinking, a cleaner context.
At this point I almost never just write it myself. I go back to the table. I chase losses, and I do it constantly and so does everyone I know.
Execution is getting cheap
Good product development used to need three things: tight alignment and fast execution with a cracked team. Execution was the expensive half, so you hired for it.
AI made execution nearly free. Now anyone can have a working app by lunch. But it also removed the pressure that forced teams to align before they built anything. The result is a graveyard of half-working apps that barely anyone uses, probably full of security holes you'll be fixing for weeks.
This shadow cost never shows up on the token meter. Everybody says the tokens were cheap and getting cheaper, but the mess is where the money actually goes: the bugs, the rework, the half-built things nobody owns.
How to not lose your shirt
The goal of this isn’t to tell you to spend less on AI. The goal of this is to make you aware that the house always win. More companies should be spending time thinking about the outcomes of all these agent runs vs. just the raw input. I’ll leave you with 2 ideas:
- Align your agents first. Good teams plan before they write code. If you let twenty of them run loose and clean up the mess later, your bill takes off. There's some interesting ideas to explore around better coordination in this article.
- Pay for outcomes. Choose AI vendors that are aligned with your incentives. You should choose vendors pricing on outcomes that are relevant to you if possible vs. raw tokens. Sierra is a great early example here for customer support and Crosby for sales contracts.
Finally, just remember at the end of the day, the Labs are selling you chips and the only thing that matters for you is what you walk out with.
