Tokenmaxxing: spend machine intelligence well

AGI claims, robot videos, and a $20 seat

We live in a bizarre period of history, and it's only going to get much more bizarre before the year is finished.

OpenAI, in a recent TIME article, claims that by year's end they will have an internal system Sam Altman would call AGI,¹ an Anthropic employee on X² casually shares another breakthrough in mathematics via AI, and China showcases³ advances in robotics that a year ago would've seemed more than five years out.

The acceleration is accelerating.

One need look no further than the incredibly powerful agentic tools offered up by OpenAI and Anthropic via Codex and Claude Code respectively.

These systems, as I've argued elsewhere, allow for types of work and automation hitherto undreamt of.

They also consume an absolutely stupendous amount of tokens.

And, more importantly, require some skill and practice to get the most out of them without running into limits or, if you're using them through the API directly, incredibly high bills.

As an agency, we are on OpenAI's Business plan and Anthropic's Team plan. Both now offer premium seats with far more usage than their standard counterparts, but let's stick with the common standard seat across both.

Through these seats, you get a certain amount of usage inside of these agentic systems. Underneath that usage are tokens.

You can think of tokens in this context as units through which machine intelligence gets consumed.

Which, back to the opening sentence for a moment, what a bizarre statement.

According to a June analysis from SemiAnalysis⁴, a $20 Claude Pro seat could represent roughly $400 a month of equivalent usage at API prices if fully consumed. Their analysis of ChatGPT was considerably higher: roughly $700 worth of API-equivalent usage for the same monthly seat price.

Some caveats. This was as of June and these things change quite a bit. Also, I use both and am a huge proponent of using both, so this is in no way me throwing shade at Anthropic.

Anyways.

With the foundation laid, let's get to the central thesis.

Tokenmaxxing

It would behoove you to maximize the amount of useful work you can get out of every token you've been given to consume.

My personal belief is that most people should be consuming much more of those tokens than they currently are and might, in fact, be in need of an upgraded seat.

Send your boss this article.

Jensen Huang⁵ reportedly said that if NVIDIA had a $500,000 engineer who wasn't consuming at least $250,000 worth of AI tokens every year, he would be deeply alarmed.

Bizarre.

But if you think about it for more than four seconds, I think you would agree.

You now have access to machine intelligence that can read your files, operate software, write and execute code, work across the web, evaluate its own work, stay on task for an incredibly long time, run on schedules and even be kicked off by triggers.⁶

With all of those capabilities, there should probably be a compelling reason for leaving most of that capacity unused.

Welcome to the token economy

For most of my career, the scarce resource in knowledge work was my time, attention and ability to sit in front of a computer and do the thing.

Sure, software made me faster, but UiPath never independently spent three hours automating a workflow while I went out for lunch.

Machine intelligence adds another productive resource to the equation.

An employee can now sit down in the morning with eight hours of their own time and mental resources while also having potentially hundreds of hours of machine work available in parallel.

I don't think we've remotely internalized how strange this is.

Again I say, bizarre.

The employee's ability to get work done is increasingly constrained by how well they can orchestrate all of that machine work and how much useful output they can extract from the tokens available to them.

Being good at your job is increasingly going to include being good at spending tokens. So, as the young people like to say, get good.

A helpful mental model

I find myself starting to think through this as an equation. This approach has worked really well when choosing a model, which reasoning or effort level, and creating, evaluating and refining skills.

By model I mean choosing between the different tiers of models each provider offers. Fable vs Opus vs Sonnet vs Haiku for Anthropic or Sol, Terra, and Luna for OpenAI (these OpenAI model names will definitely age like milk so swap them out with the latest names).

Let's define a skill as a reusable package of instructions, rules, examples, scripts and evals that teaches an agent how to perform a particular type of work.

Say I have a skill, x, that reliably produces some output, y.

Today it might look like:

skill x + opus or sol across n interactions + z total tokens → output y

And the output ends up being pretty awesome.

Great! Cool!

Now I want to know whether I can improve skill x enough that:

improved skill x + sonnet or luna + fewer n interactions + fewer z tokens → roughly the same output y

That is a very interesting optimization problem.

Somewhere in there is the cheapest combination of model, skill, and tokens that still reliably produces the output I want.

Maybe I can't get there. Maybe the frontier model really is that much better at this specific task.

Fine.

But I want to make absolutely sure I'm not spending frontier-level machine intelligence when a lower-tier model could produce the same outcome if I had built a better skill around it.

The more of my process, taste, evaluation criteria, and rules I can encode into the skill, the less work I am asking the model to figure out from scratch.

There's another variable here too: me.

A workflow that uses fewer tokens but requires me to babysit it for three hours may be considerably more expensive than one that burns more tokens and comes back finished without me.

Token minimization is not the goal.

Useful work is.

Pay attention to what the tokens are being spent on

If you use these agentic systems heavily, I highly recommend watching a full run.

Watch its reasoning, actions, and tool calls.

The agent opens a file, reads some code, forms a hypothesis, makes a change, runs the application, notices the change broke something, inspects the error, changes its approach, and runs again before completing.

Tokens well spent.

Then it reads your entire codebase before updating a button.

Maybe not.

It tries an approach you already tried last week, discovers the same failure, and backs out because neither you nor the agent ever encoded what you learned.

Now I'm annoyed.

It asks a PhD-level frontier model to perform arithmetic or makes five tool calls to do something a deterministic script could return in milliseconds.

Now we're setting money on fire.

The usage meter will happily confirm that you used 100,000 tokens.

It cannot tell you whether those tokens produced anything worthwhile.

Two people can consume 100,000 tokens and produce wildly different quantities and qualities of work.

That is why I keep coming back to some version of useful work per token.

The exact metric is less important than the idea behind it.

What did you get for the machine intelligence you consumed?

Some of your best optimization involves less AI

As part of our TOM Index, I have a method of generating hyper-personalized analysis videos for brands based on their individual findings.

The rough process looks like this:

run the TOM Index → extract a top brand from a category → run deep analysis across the Top of Model dimensions → invoke the TOM Snapshot skill → convert the analysis into a ten-beat HTML story → convert the HTML into video

Let's focus on the TOM Snapshot skill.

It ultimately produces two artifacts: a ten-beat branded HTML story and a narration transcript written for voiceover generation.

Those artifacts require judgment and taste.

We're taking dense analysis across hundreds of LLM responses and thousands of URLs and turning it into a compelling story. What do the findings mean? Which ones matter? How do they fit inside ten beats? How do you write the narration so it doesn't sound like a droning press release?

I want a strong model involved there.

But the model does not need to guess the TOM Index colors, fonts, or spacing. Those are encoded.

It does not need to invent the HTML structure. There is a template and config.

It does not need to decide how files should be assembled. A Python script does that.

The skill includes a prompt, but the prompt is only one piece of it.

It is an amalgamation of decisions I've already made, code for everything that should stay deterministic, and a deliberately small space where the model gets to exercise judgment worth spending machine intelligence on.

Encode what each run teaches you

Every run teaches you something.

Maybe the model consistently misses the same edge case. Maybe it loads too much context. Maybe one instruction sends it down a predictable dead end. Maybe you find yourself correcting the same thing for the fifth time.

Or maybe you notice a frontier model consistently understands implication and infers what you want without clear direction, while a lower-tier model does not.

Sorry your prompt sucks, man.

Jokes aside, that is a positive finding. With a bit of rigor, clear-eyed elucidation and more thorough direction, the lower-tier model may well produce the output you wanted. And for a fraction of the token cost.

So encode it.

Change the instructions. Add an example. Write a script. Add an eval. Improve retrieval. Remove an unnecessary model call.

Every subsequent run should inherit what you learned from the previous one.

This is where skills get really interesting. Over time they accumulate judgment:

• Here is how we do this
• Here is what good looks like
• Here is what failed before
• Here is what requires judgment
• Here is what code should handle
• Here is how you check the work

A good skill gives the model less figuring out to do each time it is invoked.

And as models improve, work that requires frontier intelligence today may be possible on a lower-tier model tomorrow without degrading the output.

That is an amazing economic outcome.

Same models, very different economics

Outside of working at one of these labs, everyone can buy the same intelligence.

Your agency can use the same models we do. So can your competitors.

Access does not mean, however, that you have the same AI capability.

Imagine two agencies running similar Chief of Staff agents every morning.

AgencyOne has no meaningful retrieval strategy, so its agent repeatedly pulls months of email, calendar data, old briefs and reference files into context every morning.

AgencyTwo retrieves what changed recently plus whatever historical context is relevant to that particular morning.

Both agents may produce a similar brief.

One burns dramatically more machine intelligence to get there.

Now multiply that across every employee, every agent and every repeated workflow in the company.

Those are two very different businesses.

The difference shows up in cost, margin, speed, and capacity.

And it compounds.

Every time you improve the system, you potentially get more useful work out of the next token.

So, spend the tokens

I suspect this conversation will continue to become a pressing issue throughout the rest of the year. Right now, most folks are not even aware of these harnesses or tokens. They're stuck in ChatGPT era circa 2023.

But, in the near future, the conversation will change to mostly limits. It's happening now in our agency. Why did I hit mine? When does it reset? What does the next tier of seats offer in terms of expanded token usage? Why in the world did that single Codex run use 30% of my weekly usage limit?

Fair questions. But, eventually the conversation will become much more sophisticated.

Companies already obsess over human capacity. We measure everything. Hours, salaries, utilization, rates, headcount, and software cost per employee.

Now we are adding machine intelligence to the fray as another productive resource available to that employee.

Which means we will eventually stop asking how many tokens somebody consumed and start asking what those tokens bought.

So, to wrap this all up: use your tokens and use them wisely.

Use the ridiculous amount of machine intelligence at your disposal these companies are currently willing to bundle into your monthly seat.

Give the models real work because they can more than handle it. But pay attention to what those tokens are being spent on.

Spend frontier-intelligence level tokens where frontier intelligence is required. Use cheaper models where it's not. Remove AI entirely when code has the answer.

Being good at your job is increasingly going to include being good at spending tokens of machine intelligence.

So get good, and welcome to the token economy.

Why does methodology matter if everyone can access the same AI models?

Models are broadly available. The systems around them are not. An organization that has encoded its process, rules, checks, and accumulated lessons can produce the same work with less model reasoning and less supervision, and that gap compounds every time it runs.