AI Operations12 min read
Tokenmaxxing vs Tokenleaning: Confessions of a Recovering Opus Addict
Unlimited premium-model access makes every problem look like a premium-model problem. A more durable operating practice starts with the job, the risk, and the value of a better answer.
By Dan Grams, Founder & Principal Consultant
My take
Tokenleaning means using the least expensive model that reliably clears a task-specific quality bar, while reserving premium reasoning for work where better output changes the business decision.
What matters most
- Route by task difficulty and consequence, not habit.
- Evaluate complete workflows; cheap tokens can still produce expensive rework.
- Place spending limits and quality checks at the workflow level.
- Re-test routing as models, prices, and the work itself change.
Tokenmaxxing feels productive because the meter is hidden
When the best model is always available, using it for everything feels rational. The outputs are polished, iteration is fast, and the marginal cost is easy to ignore until a plan changes, a team scales the behavior, or one successful workflow becomes fifty.
The resulting bill is not the only problem. Overpowered defaults make the system brittle because nobody has defined which tasks truly require that capability.
Route work by consequence
A product description draft and a pricing recommendation do not carry the same risk. Neither do extracting order IDs and interpreting a noisy cohort analysis. Start with the consequence of a bad answer, then consider task complexity and how easily a person can verify the output.
- Low consequence, easy to verify: small or local models may be enough.
- Moderate consequence, structured output: use a capable default with validation.
- High consequence or ambiguous reasoning: route to the strongest model and require human review.
- Repeated high-volume work: invest in tests, caching, and deterministic preprocessing before buying more intelligence.
Cost per successful outcome beats cost per token
A cheaper model that causes twice as much correction is not cheaper. A premium model that saves an executive hour may be a bargain. The unit of analysis should be the completed, acceptable business task.
Build a small evaluation set from real work. Score factuality, completeness, format compliance, reviewer time, and task success. Then route with evidence instead of model folklore.
Put a budget around the workflow
A production AI workflow should know its maximum attempts, escalation route, and acceptable latency. It should log model choice, cost, failure state, and the version of the prompt or tool chain that ran.
That instrumentation turns a surprising invoice into an operating signal. More importantly, it makes quality and cost a joint optimization problem instead of opposing slogans.
Lean is deliberate, not cheap
Tokenleaning does not mean refusing premium capability. It means earning the premium route with the value and difficulty of the task. Some work should use the best model you can access. Most work should use a system designed well enough to know the difference.