Tokenmaxxing the Weekend: Three Days with Claude Fable, and What Spending Tokens Like Fuel Taught Me About Priorities

Severin Sorensen
22 hours ago
7 min read

Optimizing AI Token Spend, while Minimizing Expense

This past weekend I conducted an experiment on myself as much as on a machine. Anthropic's new Claude Fable model had been available to me for three days, and the announcement had already come down that access would soon move out of standard subscription allowances and onto metered usage credits today. The clock was running. Rather than ration my remaining allocation cautiously, I decided to do the opposite. I resolved to spend it deliberately, completely, and on the highest-value work I could assemble, treating every token as precious fuel and flying the aircraft until the gauge read empty.

This is a refined form of tokenmaxxing: not wasting tokens as fast as you can, but the opposite. How to invest them in highest-order projects as intelligently as you can, on a fixed budget of tokens. It's an adaptation of an old Goethe saying: "As in art, the best is good enough, on budget." I want to describe both the experiment and what it revealed, because the lessons extend well beyond one weekend with one model. They reach into how leaders should think about attention, scarcity, and the allocation of any resource that matters.

The Setup: Choosing What Deserves the Best

Anyone who has worked seriously with frontier AI models learns quickly that capability is not evenly needed across tasks. A great deal of daily work, such as summarizing a document, formatting a table, or drafting a routine email, can be done perfectly well by lesser models at a fraction of the cost. The scarce and expensive resource, whether measured in tokens, dollars, or rate limits, should be reserved for the work where the marginal intelligence of the best available model actually changes the outcome.

So before I spent a single token, I asked the question every operator should ask of an expensive asset: what work is worthy of this? Or as Clayton Christensen might have said, what are today's most critical and crucial 'jobs to be done.' I settled on three substantial projects, each chosen because it demanded sustained reasoning over many hours, synthesis across large bodies of material, and creative extension beyond what I had done before.

The first was curriculum development turning a past AI workshop into an online AI course, a full program of 78 video lessons that required coherent instructional design across the entire arc of the course, not merely a stack of disconnected scripts. The second project was a new book project, which meant reviewing my past manuscripts, my serial publications, new research, and roughly two years of accumulated musings, asking the model to hold that entire corpus in mind while evaluating what the new work should become. The third was the most speculative and the most fun: pondering and writing code for the needs of 2030. The decade's end is closer than it feels, and I wanted to press the model for everything it could do now, drawing on my past work but pushing it, ever so gently, to stretch further. I was pleasantly surprised by what came back.

Along the way I gave the model supporting roles as well. It played IT consultant to troubleshoot technical issues as they arose. It helped me develop concepts for teaching first principles of AI to executives. Like in my AI Workshops, I had several AI plates spinning at once, and I kept my system connected through the weekend so the work could continue throughout the night without me, compressing what would otherwise have been weeks of effort into a weekend.

The Ceiling: What Happens When the Fuel Runs Out

The experiment had a second purpose beyond the work itself. I wanted to see how quickly a heavy, sustained, multi-project workload would exhaust the Fable allocation on my Anthropic $ 200-per-month Max plan. The answer arrived before the weekend was completed: it maxed out with about 85% of the work completed.

What happened next is the part worth dwelling on. When my access dropped back to the previous flagship model (Opus 4.8), the productivity of the whole operation fell noticeably, and not because the ceiling of raw capability had lowered. The most disruptive symptom was continuity. Like a light switch flipping off, the system forgot where it was in the process, losing the thread of work that had been proceeding smoothly for hours. Let me be fair to the older model: it remains excellent, and six months ago I would have described it as remarkable. But once you have worked at the new level with Fable, the previous one occupies second place in your mind, and you feel the difference in every exchange. It's like putting your 2nd string out on the playing field when you know your better players are benched by a penalty. Or another metaphor: losing access to Fable was like returning to a smaller monitor after a week on a large one, where nothing has failed, yet the working environment feels constrained in ways one notices instantly.

I offer this less as a complaint than as an observation about the psychology and economics of capability upgrades. Each new tier of intelligence resets the baseline of what feels acceptable, and that reset has consequences for how we as managers budget, how we plan, and how we feel about the tools we thought we loved last month.

The Deeper Lesson: Scarcity Is a Teacher

Another lesson. What surprised me most about the weekend was the discipline that deliverable-focused tokenmaxxing forced upon me, namely the ruthless prioritization of what deserved the best model's attention, turned out to be a better use of my own judgment than almost anything else I did. We have a role with AI, and it serves up as acumen, prioritization, taste, and discernment of 'what's next.'

When a resource is abundant, we spend it carelessly, as if it had no cost, and when it becomes scarce and expiring, we suddenly discover our inner economist. I found myself asking, for every task that crossed my mind, whether it was worthy of the premium model or belonged with a lesser one, and this understanding did more to clarify my actual priorities for AI builds than any planning exercise I have run this year.

The parallel to leadership is direct. We prioritize customers because not all accounts warrant the same investment of our finest attention. We prioritize conversations, giving our best coaching hours to the team members where development will compound. And, most importantly, we prioritize time with the people who matter most, because the hours of a life are the original metered resource, allocated to us without a published rate card and without the option to purchase more. If I am willing to architect an entire weekend around extracting maximum value from a temporary allocation of machine intelligence, I ought to bring at least that much intentionality to how I allocate myself.

The Economics: Building a Roster, Not Buying a Star

Several observations surfaced for me on the same weekend as the announcement of new pricing policies, tiers, and API throughput arrangements for the new model, which sharpened the question from an experiment into a budgeting decision.

I am not yet convinced that I will increase my spend. My monthly investment in AI across providers is already substantial, and the arrival of a more capable and more expensive tier does not automatically justify a larger budget. What it justifies is a reallocation. The right mental model, I have come to believe, is that of a sports team coach. A coach does not play the star in every minute of every game. Each player is deployed for the situations where their particular utility is highest, and a place on the bench reflects strategy rather than any slight to the player.

Applied to AI, this means routing work deliberately. The frontier model gets the problems where its marginal intelligence changes the outcome: deep synthesis, long-horizon reasoning, creative extension of hard problems. Capable mid-tier models carry the substantial middle of the workload. Commodity models handle the routine. Other providers' offerings fill the roles where they hold genuine comparative advantage. The objective is not to own the most expensive roster in the league but to win games at a sustainable payroll.

For enterprises, this discipline is arriving whether leaders are ready or not. As frontier access shifts toward metered consumption, token allocation becomes a management question rather than an IT line item, and the organizations that develop the habit of asking "what work is worthy of our best intelligence" will hold a quiet but compounding advantage over those that treat all AI usage as interchangeable.

What I Would Tell a Fellow Executive

If you are weighing how to approach this new generation of models, my weekend suggests a few practical conclusions.

First, run your own tokenmaxxing experiment. Take a bounded window of access and spend it entirely on your highest-value problems. You will learn more about the model's true ceiling, and about your own priorities, than any benchmark or review can teach you.
Second, prepare for the fallback. Whatever tier you operate on, there will be moments when you drop to a lesser model mid-process. Structure your work so that context is preserved outside the conversation, in documents and artifacts the next session can pick up, because continuity is where the productivity losses concentrate.
Third, budget by allocation rather than by aspiration. The useful question is not whether the frontier model is worth the price in the abstract, but rather which specific work in your portfolio justifies frontier pricing, and how the remainder should be routed to the rest of the roster.
Fourth, and finally, let the exercise instruct you beyond the technology. The same scarcity logic that governs tokens governs your calendar, your coaching hours, and your evenings at home. The machine taught me nothing this weekend that the wisest people in my life had not already tried to teach me. It simply presented the lesson with a usage meter attached, and there is something clarifying about watching a gauge fall in real time.

The fuel always runs out, and the only question that remains at the end is whether it was spent on what mattered.

I would be interested to hear how others are approaching this transition. Are you increasing spend for frontier access, reallocating across providers, or waiting to see how the pricing settles? The comments are open, and so am I.