$500k on AI tokens. 80% was waste.

We wasted 80% of our AI spend. Here are the 5 moves that fixed it.

Amos Bar Joseph July 16, 2026

I’m Amos Bar Joseph**, co-founder of Swan, the first Autonomous Business OS. At Swan, we’re building what we call the Autonomous Business: a company that scales to $10M ARR per employee with no bloat, no assembly lines, no Cog Culture. Just humans in their zone of genius, amplified by AI agents.*** I write The Autonomous Age to share contrarian insights from that journey, on GTM, leadership, and the future of work. If you want to understand how GTM evolves beyond playbooks and assembly lines, this is where the story unfolds. Connect with me on Linkedin or X. If that's not the game you're playing, reply to unsubscribe.

We spent $500k on AI tokens in the last 4 months (we're a 6 person team).. here are 5 ways you won't go bankrupt doing the same:

For context, at Swan AI** we're building an autonomous business.

A business designed from the ground up to scale with intelligence, not headcount.

Our goal is to get to $10M ARR per employee.

and that means spending a sh*t ton on tokens to get there.

So for us, the answer can't be "use less AI."

Instead, we focus on "reducing AI waste"..

5 simple moves that dropped our AI waste 80%.. and you can run every one of them in the next 30 days (the last one is the hardest but biggest):

• The "Big Spender" Waste.

Usually there's this one guy who sent claude on a mission to build a rocket ship to the moon last week and his agent is still running since then. But you can't cut what you can't see.

Turn on the usage view.. by person, by team, by model. most leaders have never once looked.

• The "One Chat to Rule Them All" waste.

You open the Claude tab, it's already open on a long convo you had last night, but you just have this one quick question - so you quickly write it down and waste 500k tokens on "what's the meaning of anthropic?"...

One endless "everything" chat drags monday's context into friday's question, and you pay for the whole pile.Remember - one task, one thread.

• The "Overkill" waste.

I get it. You don't want to read the article your boss sent you yesterday. And you want to feel like you're "AI-native", so you send Opus 4.8 on High Effort to summarize an article about "5 different ways to appear on ChatGPT".

Match the model to the job. save the frontier for frontier problems. Start with the simpler models, and if they fail you - move up the chain.

• The "Repetition" waste.

Pasting the same 30-page doc into a fresh chat 40 times a week bloats your bill more than the burgers you ate at your uncle's 4th of July's BBQ.

Load text-heavy files into a project once, ask it forever, pay for it once (anthropic caches it).

• The "Frontier" waste.

This is the biggest one, and the one we just fixed at Swan. Our agent ran on Sonnet + Opus, $100k a month, on repeat. then finally Open Source models caught up - MiniMax and GLM dropped, 80% cheaper with quality on par with Sonnet.

So now only the hard stuff goes to the frontier's Opus, and the boring 90% goes to the cheap model that does it just as well. this is the BIGGEST lever you can pull, but requires some technical setup.

Remember, it’s not about capping everyone.

Just focus on driving the waste to zero.

What's your best practice on reducing token spend?

*[getswan.com

](https://getswan.com/?utm_source=theautonomousage&utm_medium=referral&utm_campaign=500k-on-ai-tokens-80-was-waste)

AI spend isn't the problem, invisible waste is, so we started watching who spends what, kept every task in its own thread, matched each job to the cheapest model that could still do it, and cut our bill 80% in 30 days without touching anyone's access.

*The waste audit

• Find the agent still running from last week. Kill it today.

• New deal, new thread. Every time.

• Move templated outreach to a lighter model this week.

• Drop your deck and case studies into one project. Stop re-pasting.

• Best model for your stuck deal. Cheap model for the rest.

What AI task still gets your best model, out of habit?

Scale with intelligence, not headcount, and know where every token goes.

-Amos

Read more.