# My Tokens Cost More Than My Salary

**Rx Recursive Digest · Issue 07** · A reflection by Jack Wang · Issued 24 Oct 2026
10 min read · digest.reinventionx.app/digests/issue-7-tokens-cost-more-than-salary

Recently, I realised that my monthly AI token cost exceeded my salary. That felt worth pausing over. The number started an audit of the work, the work turned the question back on me, and by the end I was asking a better one.

`// a reflection`

This one is a reflection rather than a briefing. Each section picks up the question the one
before it raised, one level further down, and the last one comes back to where it started.

`depth 00`

**Recently, I realised that my monthly AI token cost exceeded my salary. That felt worth pausing over.**

> A qualification before anyone in finance starts looking for me: our access is covered through the enterprise. The number on my dashboard is an estimated value based on the list price, not necessarily what the company is paying. Those are different things, and [Claude Code’s documentation makes that distinction too](https://code.claude.com/docs/en/costs). So yes, the title is doing a bit of heavy lifting.

But even with that understood, seeing the number did crystallise a few things.

There was excitement. A slightly smug moment. Some nervousness. And, eventually, a more useful way of looking at it, which is what I want to share here.

---

`↳ depth 01`

## It started with the setup

The journey to that number began with me mucking about with [Herdr](https://github.com/herdrdev/herdr), a terminal-based environment for working with coding agents. This is not a piece about Herdr, although it is a good tool.

Different sessions, different projects, agents sending messages between them and coordinating tasks. I was reading the agents’ exchanges, adjusting the setup and enjoying the feeling of having a working command centre.

While I was fiddling with it, I realised my setup was missing something: a proper agent-operations view. I could follow the work, but I could not easily see the usage behind it.

So, partly inspired by another tool, [Mole](https://github.com/tw93/mole), I built one.

Live token usage. Minute by minute. By project, task and model.

![The agentop terminal dashboard. Today’s estimated spend is about $1,348, across nine Herdr panes in two masked projects, with subagent counts, context bars and the cost of each pane.](/images/issue-7/agentop.png)

*agentop, the agent-operations view I built for Herdr, with the project names masked. The same agents it now watches built it. Every figure on it is an estimate at list price.*

I was rather pleased with it. I finally had a much clearer view of what was going on across the sessions, rather than a collection of counters I had never quite trusted.

Then I started looking at the numbers.

---

`↳ depth 02`

## The imaginary leaderboard

Initially, I was proud.

Not because I had compared the usage with anything sensible. I had not got that far.

I was watching all this capability being put to work. Projects moving. Agents working through problems. Things getting built.

Somewhere in my head, I had entered an imaginary token-usage leaderboard. And apparently I was doing quite well.

Speaking as someone who refuses to admit he is competitive...

Then, at some point, it landed. That is more than I earn.

I went from enjoying the number to wondering what it meant. Was this impressive? Wasteful? Was I getting carried away? And why had comparing it with my salary suddenly made it feel different?

Until then, I had not needed to pay much attention to consumption. Access was there, the work was getting done, and I had mostly used the most capable models available.

So I started auditing the projects.

---

`↳ depth 03`

## What would not shipping cost?

Which projects had consumed the most? What had the agents been doing? What had actually shipped?

Some of the usage was easy to question. I had defaulted to the strongest models rather than choosing them for a specific reason. There were workflows to tighten and tasks that did not need that level of horsepower.

But looking through the work also changed the comparison.

What would it otherwise have taken to deliver these things? Not just development time, but research into new domains, industry expertise, finding capacity, agreeing priorities, coordinating people and waiting for a slot.

More importantly, what would have happened if they had not been delivered at all?

A deal at risk. A deadline missed. An escalation. Another commitment someone had to explain to a stakeholder. Trust becoming a little harder to maintain.

These were not hobby projects I could abandon once the interesting part was over. They were things people wanted because they helped them do their jobs better.

And there was a snowball effect.

1. You ship something for someone.
2. They use it.
3. It helps.
4. They come back with another problem, or show it to a colleague who has one of their own.

*↩ and round again, a little bigger each time*

The expectations grow — not because everyone wants to join in the token-maxing, but because the work genuinely matters to them, delivered at the speed and quality that make a difference to their stakeholders and the business.

I doubt most of them have any idea how many tokens are involved.

They just know whether something arrived when they needed it, and whether it helped.

> The cost of building was visible. **The consequences of not building needed to be considered as well.**

---

`↳ depth 04`

## Why me?

Comparing the usage with my salary missed most of the context. But it did leave me with an awkward question. If that much capability was running through me, what was I actually adding?

Am I simply someone who picked up a few AI tools early, got on with it and became reasonably good at using them? Or is there something more specific I am bringing to the work?

Looking back through the projects, I could see two things happening.

First, some of what had impressed me in the early deliveries was already becoming ordinary. The HTML, the static websites, the standard vibe-coded interfaces. Things I had been pleased to get working were becoming things many more people could produce.

That is not a criticism of those projects. They did their job when they were needed. But being able to generate them was never going to remain much of a distinction.

The more interesting part was what made each one right for the audience it was built for.

The narrative. The vision behind it. The sequence in which an idea needed to unfold. A particular interaction that helped someone understand a difficult concept, or content that reflected the client’s actual situation rather than sounding plausible for any business.

Those details were not finishing touches. In many cases, they were the reason the thing was worth building.

AI helped me develop them, challenge them and make them work. I would not pretend every good idea was mine alone. But neither were they simply sitting inside the model waiting for anyone to retrieve them. They came out of the context I brought, the questions we worked through and the choices made along the way.

Then there was the second change: what happened after delivery.

What people used. What needed explaining. What missed the point. What I had to fix. Each project gave me feedback I could take into the next one. Not just a better prompt, but a clearer understanding of what mattered.

> **The tools** — were making more of the implementation commonplace.
> **The experience** — of applying them was making my understanding **more specific.**

That does not make me irreplaceable. It does mean I am not starting from scratch each time I open a terminal.

So no, I do not think I am just another person who happens to know a few AI tools. Knowing the tools is part of it. So is understanding the client, shaping the experience and learning from what happens when someone actually uses it.

The pride came back, but it was less about the number now. I could see more clearly what the consumption was enabling, and what I was contributing.

I could also see the responsibility that came with it. Once people rely on what you ship, quality, maintenance and setting realistic expectations become part of the work, not things to sort out after the exciting bit.

That is closer to what the job means to me: understand the problem, build something that helps and get it into the hands of the people who need it. Then stay accountable for whether it works.

---

`↳ depth 05`

## What carries forward

The salary comparison had changed how I saw the whole thing.

For now, I have unusually generous access to a capability I am still working out how to use. That access is worth a lot to me, not just for what I can deliver today, but because working this way keeps changing how I approach the next problem.

Not every token contributes to that. Some are wasted. Some produce work I discard. And some become experience I can use again.

A failed approach can make the next decision better, provided I understand why it failed rather than just asking an agent to try again.

**The dashboard can price the usage. It cannot tell me how much of that experience I will carry into the next project.**

That is part of the reward I had not been accounting for.

None of this replaces being paid fairly, and I am not pretending it does. But having room to explore, build and develop judgement in an area I care about is worth something in its own right.

I am enjoying that access. Mostly. There is still a little nervousness mixed in, partly because I have got very used to working this way, and generous access is not guaranteed to stay generous.

---

`↳ depth 06`

## A machine in one hand

Watching Jensen Huang’s recent GTC Berlin keynote put all this into a wider context.

He held up a compact machine, comparing its stated one-petaflop performance and 128GB of memory with the DGX-1 he delivered to OpenAI in 2016, a decade ago ([NVIDIA’s keynote account](https://blogs.nvidia.com/blog/gtc-berlin-2026-keynote/)).

**2016 · San Francisco**

![Jensen Huang opens the first DGX-1 on a table at OpenAI in San Francisco while Elon Musk watches with his arms folded.](/images/issue-7/dgx-1-2016.jpg)

**2026 · Berlin**

![Jensen Huang on stage at GTC Berlin in front of a giant screen showing an RTX Spark laptop, its chip and two press quotes about it.](/images/issue-7/gtc-berlin-2026.jpg)

*In August 2016 Huang delivered the first DGX-1 to OpenAI by hand, with Elon Musk, then one of its co-chairs, looking on. In Berlin in October 2026 he put RTX Spark beside it. “I can hold this like this,” he said, “and DGX-1 took two people to carry upstairs.”*

It was a striking image of how much the form factor had changed, even if the two machines are not identical in every technical respect.

The part that stayed with me was the question of access. What happens when capabilities previously associated with specialist teams and serious infrastructure become much easier for an individual to work with?

In my own small corner of that change, I can already feel my assumptions moving. Problems I would once have assumed needed a team, I now start investigating myself. Often the first question is no longer who to involve, but whether I need to involve anyone yet. And when a team is needed, I can get further before making that call and turn up with something concrete to discuss.

Huang’s framing of AI infrastructure as factories producing tokens also gave a name to something I could now see on my own dashboard.

I do not know where token prices will settle. A cheaper unit of computation does not guarantee a smaller bill if I keep finding more worthwhile work for it to do.

But I can imagine a working environment in which access to tokens feels as ordinary as access to data does today. Not free, not without limits, and not equally available to everyone, but a normal part of how work gets done.

If that is the direction of travel, learning to use it well matters more than proving I can consume a lot of it. Although, to be fair, you do have to consume quite a lot of it to find out what it can do.

Which is the uncomfortable bit. Andrej Karpathy recently argued that understanding of AI capability is diverging, shaped by access, use and the problems people attempt ([his October reflection](https://x.com/karpathy/status/2109361546505966046)).

**Figure 1 · Karpathy’s funnel, drawn to scale — Three zooms from everyone to the frontier**

Karpathy’s own estimates of who has seen what, from his post of 11 October 2026. Each frame is a grid of squares, and one square in it is the whole of the next frame. Open the ringed square, or use the steps above the grid.

*Interactive figure: Everyone (~8 billion), then Frontier, math & code (~20 million), then Inside the labs (~5,000), each frame one square of the frame before it.*

| Tier | People | Share |
| --- | --- | --- |
| Barely in contact with LLMs at all | ~6 billion | ~75% |
| Casual, infrequent free-tier use | 1–2 billion | ~20% |
| Professional frontier use, math & code | ~20 million | 0.2% |
| Internal access to frontier systems | ~5,000 | ~0.00006% |

- *For scale, everyone:* Microsoft estimates that 18.8% of the world’s working-age population used AI in June 2026. [Microsoft AI Diffusion Report](https://blogs.microsoft.com/on-the-issues/2026/09/21/the-continued-state-of-global-ai-diffusion-in-2026/)
- *For scale, frontier, math & code:* SlashData counted about 36.5 million professional developers worldwide in early 2025. [SlashData](https://www.slashdata.co/post/global-developer-population-trends-2025-how-many-developers-are-there)
- *For scale, inside the labs:* No public count exists. This tier is his estimate alone.

*Karpathy calls these numbers “napkin math”, and the chart draws them as he gave them, not corrected. His shares imply a world of about 8 billion, so the first frame is 400 squares of 20 million. The casual tier is a range of 1 to 2 billion, and the outlined squares are the rest of that range. The post is itself a second pass. **He quoted his own post from April, about two groups talking past each other, and found that six months on the gap had become a funnel.***

I can see that gap opening around me. The people who are building and shipping are learning things that are hard to pick up any other way.

That is why I keep nudging my engineers to build and ship. Reading about it only gets you so far. Getting stuck in is the only way I know to catch up.

---

`↩ back to depth 00`

## A different comparison

I still want to reduce waste. Now that I can see the usage, I have fewer excuses not to.

But I do not want to optimise only for the smallest number on the dashboard. And I do not want the largest number to become a badge of honour either.

So I am back to the questions from the audit, just asked a bit more honestly: what shipped, did it help, what did it cost to deliver and keep running, and what do I understand better as a result?

The comparison with my salary got my attention. Following the work told me far more.

Some of the return belongs to the business: things delivered, commitments met, opportunities pursued.

Some of it stays with me: experience, judgement and a different sense of what I can attempt next.

**I started by asking why my tokens cost more than my salary. I am now more interested in whether I am making good use of the capability, and the opportunity, I have been given.**
