Imagine sending a sticky note by international priority courier. Overnight service, tracked across three borders, signature on delivery, for a slip of paper that says “yes, approved.” Nobody does that with mail. Almost everybody does it with AI.
That is the Token Tax: the surcharge you pay every time a trivial task rides the most expensive model on your bill. It never appears as a line item. It appears as an invoice that keeps climbing while the work being done stays exactly as simple as it was.
Look at the menu. GPT-5.5 Pro and Claude Fable 5 sit at the top of the market. GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro hold the middle. DeepSeek V4 Pro undercuts them all. The per-token price across that list doesn’t vary by percentages; it varies by multiples. Same task. Same output. Wildly different price.
I learned this hands-on, building ZAMARA, my multi-model agent system. The core design decision wasn’t picking the model; it was refusing to answer that question once, globally. ZAMARA routes each task to the right-sized model: extraction and classification go to the cheap, fast tier; multi-step reasoning where a wrong answer is expensive earns the premium call. This isn’t just my house experiment. UC Berkeley’s RouteLLM research is built on exactly this per-query routing decision.
How the tax gets levied
Through defaults. During the pilot, someone asks “which model is best?”, picks the top of the leaderboard, and that choice hardens into the org-wide standard nobody re-litigates. That’s the observation. My inference: “always use the best model” isn’t a quality strategy. It’s the absence of one. A flat rule doing a graded job. Deciding which model gets which task is a routing decision, and routing is governance: a smaller sibling of the judgment layer (opens in a new tab) question of which tasks deserve a human at all.
How to stop paying it
- Audit the flow. Break your token spend down by task type, not by team. Find out what the premium model is actually doing all day.
- Grade the tasks. Sort by stakes and complexity, not prestige. Most volume is sticky notes; a minority is contracts.
- Default down, escalate up. Route to the cheapest model that clears the quality bar, and reserve the frontier tier for tasks that earn it.
A premium-everything default is Stage 2 (opens in a new tab) behavior showing up on the cost line: spending for comfort instead of designing for fit.
What to take into the room: two questions. What is our default model, and who decided the sticky note needed priority mail? If nobody can answer the second, you’re paying the Token Tax, and the refund is one routing decision away.