My Personal Tokenmaxxing Stack

Thomas Prommer Technology Executive & AI Strategy Consultant Connect on LinkedIn

Published: April 16, 2026

Updated: April 17, 2026

Two Tokenmaxxings, One Word

Silicon Valley has spent the last week arguing about tokenmaxxing. TechCrunch ran Reid Hoffman weighing in. WSJ covered Meta shutting down an internal leaderboard that ranked employees by their AI token consumption. Axios reported Salesforce pitching outcome-based metrics as the alternative. Forbes asked if the whole thing is a cult. Inc called it the controversial new AI productivity metric. The debate is real and, at the corporate level, mostly deserved — ranking people on tokens consumed is a classic category error. Engineers at Meta reportedly built bots to burn tokens in loops. That kind of metric gaming is the oldest story in software management.

But there is another tokenmaxxing that the news cycle is flattening. Before the term became Silicon Valley shorthand for gameable corporate theater, it had a quieter meaning in developer communities: actively optimizing how you extract value from multiple fixed-price AI subscriptions. That's what I actually do. I tokenmax at a desk of one. No leaderboard is ranking me. Nothing I do inflates anyone's KPI. And it saves me roughly a thousand dollars a month compared to routing the same work through pay-per-call APIs.

This post is the stack. What I pay for, what each subscription does, how I route work between them, and what I would drop if I had to cut half the budget. For the broader definition of tokenmaxxing and the corporate debate, see the tokenmaxxing explainer I wrote on ctaio.dev. This page is the personal version.

The Stack: What I Actually Pay For

As of April 2026, my monthly AI coding spend is $547 spread across six subscriptions. Here is the full inventory with what each does that the others don't.

The $50 on fal.ai is technically media, not LLM, but it lives in the same spreadsheet and serves the same purpose: a fixed-price ceiling on a variable-cost activity. The total excludes ChatGPT Plus, which I dropped when Claude Max 20x started replacing it for every task I had been paying OpenAI for. That cancellation is its own small case study in why you should review the stack monthly.

The Math: Why Stacking Beats Single-Subscription

The simplest argument against stacking is that six subscriptions cost more than one. On paper, yes — $547 is more than $200. But that comparison ignores the thing the whole exercise is built around: rate limits and per-task pricing efficiency.

A single Claude Max 20x account covers roughly 6-8 hours of heavy agentic coding before you start hitting meaningful rate limits. I work 8-10 hours daily, 5-6 days a week. Without the second Max account, every overflow hour would route to Sonnet 4.6 API at $3 input / $15 output per million tokens. A typical hour of agentic coding consumes 60K-120K tokens. At 2 hours of daily overflow, that's 120K-240K tokens × 22 working days = 2.6M-5.3M tokens a month. Even at an 80/20 input/output split, that's $250-$400/month in API overflow — more than the $200 I pay for the second subscription.

z.ai and Kimi work on the same logic from the other direction. Batch classification, document summarization, and cheap generation do not need Claude. They need a model with a long context window and a low per-token price. A $15/month z.ai subscription handles enough of my daily batch work that equivalent usage through the OpenAI API would cost $80-120/month. The same math applies to Kimi for long-context document tasks.

The remaining subscriptions are qualitatively different. Copilot at $10/month is the cheapest flow-state tool in computing — inline completion at <50ms latency, which no API-based setup can replicate without infrastructure work. OpenRouter at ~$60/month variable is the optionality layer: I do not need a seventh subscription if any given month produces one weird task that wants a specific model. fal.ai at $50/month is bounded media spend.

Notion as the Management Layer — Kept Small on Purpose

Every article about AI subscription stacking eventually pitches an elaborate tracking system. Dashboards, cost-per-token instrumentation, per-project attribution. I tried that. It takes longer to maintain than the subscriptions save, and it produces the wrong kind of data — precise numbers about last month when what I needed was a rough yes/no about whether each tool is still earning its keep.

What I actually use is a Notion table with four columns: tool name, monthly cost, primary use case, and a rolling weekly checkbox. Every Friday I spend five minutes marking which subscriptions got used that week. If a subscription goes two consecutive weeks unchecked, it gets flagged. Three weeks, I cancel. The friction is deliberate — if I made it easier to track, I would spend more time tracking and less time deciding. The whole point is that subscriptions compound; the five-minute review keeps compounding from turning into subscription bloat.

The secondary purpose of the Notion table is that it forces me to name the primary use case per tool. If I cannot articulate why Kimi exists in the stack in one sentence, Kimi probably shouldn't be in the stack. That framing alone has killed three subscriptions in the last year.

Honest Regrets: What I'd Drop, What I Got Wrong

The honest section. Kimi is on thin ice. It overlaps with z.ai for most of what I use it for, and the Chinese-market content use case that originally justified it is smaller than I projected. If I had to cut one subscription today, it would be Kimi, and z.ai would absorb its workload without a noticeable drop in quality.

The second Claude Max is the subscription I most often second-guess. The math works at my current usage, but the math would break if I dropped below six hours of daily Claude Code work — which could happen if I shift to more writing or more meetings in a given month. The right rule would be to cancel the secondary account the month I anticipate the shift, not to pay for two Max plans out of habit. I have not always gotten that timing right.

The biggest mistake in the history of this stack was keeping ChatGPT Plus for four months after Claude Max 20x made it redundant. I paid $80 for nothing because canceling felt like a bigger decision than it was. The lesson, which I now apply more aggressively, is that subscription cancellations should be easier than subscription additions — so if I am ambivalent about a tool, the default should be to drop it and re-add it if I miss it.

What the News Cycle Misses

The corporate tokenmaxxing debate is legitimate on its own terms — ranking employees on tokens consumed is a bad metric, and the critics (Jon Chu, the Salesforce team, most of the engineers quoted anonymously in the coverage) are right about the specific claim. But the debate keeps framing all token maximization as theater, which misses the narrower practice it borrowed its name from.

The personal version works precisely because there is no leaderboard. The incentive is pure cost optimization: stack subscriptions to the point where the marginal subscription costs less than the equivalent API usage it replaces, then stop. There is nothing to game because no one is watching. The fact that the word "tokenmaxxing" got pulled into a corporate management controversy is a branding accident, not a reflection of the practice.

The practical lesson is almost the opposite of the one the corporate debate is drawing. At the corporate level, stop tracking tokens and start tracking outcomes. At the personal level, keep tracking your subscriptions — but track them against your own costs, not against anyone else's metric. That is the tokenmaxxing that works.

FAQ