Token Cascade vs Raw Token Consumption

When teams first start tracking AI usage, they reach for the most obvious number: how many tokens did we burn? It feels like a productivity signal. It isn't. Two operators can consume the same number of tokens and have wildly different efficiency — one compounding signal, the other just spending. Raw token consumption measures volume. It tells you how much you spent, not what you bought.

The metric that actually matters is token cascade efficiency — yield (Υ). It measures the architecture of your token flow, not the volume. This post explains the difference and why volume-based tracking misleads.

What raw token count measures

Raw token count is the sum of every token that crossed the wire — input plus output plus cache-read plus cache-write. It's a billing number. It tells you what you owe the API provider. It does not tell you whether those tokens produced anything useful.

Consider two operators who each consumed 2 million tokens in a week. Operator A sent 1.8M of those as fresh input, reused almost no cache, and got 200K of output. Operator B sent 400K as fresh input, reused 1.2M from cache, and got 400K of output. By raw token count they're identical. By efficiency they're opposites.

What cascade efficiency measures

Yield (Υ) captures the shape of the cascade — how well your tokens compound — in a single number:

Υ = (cache_read × output) / input²

The formula rewards three behaviors that define an efficient operator:

• Reusing cached context — cache_read is multiplied, so operators who build on prior turns instead of re-explaining score higher.
• Producing output — output is multiplied, so operators who extract real work from the model score higher.
• Keeping fresh input lean — input is squared in the denominator, so operators who flood the model with fresh context are penalized.

Raw token count rewards none of these. It only rewards spending. That's the core difference: volume measures how much you spent; cascade efficiency measures what you got for it.

Why input is squared

The input² term is the part that surprises people, and it's the part that does the most work. Fresh input is the expensive resource in the cascade — it costs compute, it adds noise, and it doesn't compound. Squaring it in the denominator means the penalty for burning input grows super-linearly:

This is deliberate. Operators who flood the model with fresh context — pasting entire files, repeating instructions, starting every session from scratch — are the ones whose cascades burn instead of compound. The input² term makes that behavior visible in the number.

The two operators, scored

Back to Operator A and Operator B — both at 2M tokens consumed. Here's what yield says about each:

Same volume. Opposite efficiency. Raw token count can't tell them apart. Yield can. That's why leverage and yield — not token count — are the metrics that define a skilled AI operator.

When high token usage is fine

High token usage isn't inherently bad. An operator who consumes 5M tokens but reuses 4M from cache, sends 500K as fresh input, and produces 500K of output has a high yield — they're compounding, not burning. The volume is high because the work is substantial, not because the cascade is wasteful.

The failure mode is high token usage with low cache reuse and low output — burning tokens without compounding signal. That's tokenmaxxing, and yield flags it immediately. The number doesn't care how much you spent; it cares whether your spending compounded.