On Feb 6, 2026, agents used more tokens than humans for the first time. @PeterJ_Walker of OpenRouter doesn’t think that will ever reverse.
6T tokens a week in January. 190T last week.
“If you are not pricing or building for agents, I think this is a good wake-up call.” 👇
Quick test: what share of your usage is agents today? More from the After Tokens stage, and full recordings, coming.
1
32
“Close one door and they try another.”
@gareth_w’s line is the one every AI team nods along to. But look at how the doors work. Almost all of them open for the same reason: compute gets spent before anything checks whether it’ll be paid for.
Claire’s post hit home. We’ve faced the same at Atlassian. Sharing some of the key actions we took below.
With AI products, abuse isn't just annoying, it's expensive. Every LLM call costs real money and for small businesses this can be absolutely crippling.
We got hit with stolen cards, fake trials and credit-burners.
This is what helped (none of it is a silver bullet, it's a constant battle):
1. Make free trials cost something. Require a card and pre-auth it up front. Even then, we saw stolen cards and fraudsters paying just to get in the door.
2. Only let your real app talk to your backend. Verify requests actually come from your client, not just that they carry valid credentials.
3. Allow only the models you offer. Block everything else by default. Fraudsters want the latest and greatest.
4. Set hard limits per account and per org, with no overage until payment clears. It feels wrong and counterintuitive to start from "I literally don't trust you", but when needs must.
5. Bill sooner. If you charge in arrears or give a grace period on failed payments, abusers sign up with cards that won't clear, burn credits fast and vanish.
6. Use fraud tooling. Your payment provider's rules, like Stripe Radar, can flag and block suspicious accounts.
7. Watch costs daily and act fast. Suspend bad accounts and be ready to pause sign-ups. At one point we temporarily paused all new sign-ups and only sold to existing customers.
8. Restrict free trials to verified business domains. Block disposable and free email domains, and cap how many trials one domain can start.
Biggest lesson: abusers adapt crazy fast. Close one door and they try another. It's a shitty game of whac-a-mole, so plan for ongoing defence, not one-off fixes.
1
73
Then close the side doors. One AI platform estimated about $1M a year in leakage. The number dropped sharply once they fixed a marketplace integration that never wrote usage to the ledger, after 32,000 trial accounts had walked through it.
Every path to the GPU needs the same entitlement check.
1
14
If your enforcement reads a cached lock flag, the cache TTL is your fraud budget, so write that number down.
Billing tells you what happened. Stigg decides what’s allowed, before the GPU spends a cent.
The full playbook from @DorSasson →
stigg.io/blog-posts/token-fr…
19
"Agents don't eat, they don't sleep, they will eat through your token budget. This changes the rules of the game."
Jonah Cohen (@OpenAI) at After Tokens.
Then @databricks, @glean, @mixpanel, @Navan, @webflow, @qlik, @Intuit, and @temporalio took it from there. That's a wrap. 💚
1
139
Consumer growth panel with @_aliabbasrizvi from @GammaApp now at After Tokens: With humans, we know we have peak times, holidays, and such. With agents, they are unpredictable, and the best part - you can create them yourself.
4
99
Good morning from #AfterTokens, 1,000+ registered. @PeterJ_Walker (OpenRouter) set the tone:
"I wouldn't say we are in an AI bubble, but we might be in the middle of a financial mismatch."
Not even lunch yet. Next up: Vercel, Docker, PostHog, OpenAI, Databricks and more.
5
128
Stigg retweeted
Good morning San Fransisco. Final rehearsal for my keynote.
About to hit 1000 registrations. Yep - three zeros. 99% of which organic. AI builders want better quality and hard truths.
After Tokens is officially the most important AI Economy Infra event.
See you there!
This works for a PLG motion with credit top-ups and for complex enterprise contracts alike.
You can change your monetization as fast as you ship, without the engineering overhead.
Full announcement on our blog 👉 stigg.io/blog-posts/stigg-ai…
38
Almost every team builds rate limiting and billing as two separate systems. Different owners, different code paths, different problems entirely.
AI collapses them into one.
Here's what OpenAI learned rebuilding the system that decides, in real time, whether to serve your request. 🧵
2
1
265
And the closing warning: don't over-engineer for every imaginable future.
Name the knobs you'll actually need in the next 6-12 months. Build those with correctness first.
"Watch out for those trapdoors. That's where you ask the hard questions."
stigg.io/blog-posts/inside-o…
1
15






