Token Scale
From module T1 “Token Awareness”, here to try out. Four instruments, everything counted in tokens and multiples — not a single amount of money. Tokens are a unit and stay comparable, prices change.
What is a million tokens?
All the learning material in T-AI-L comes to 1,149,067 tokens across 1,157 files. A million is the share the dark marker shows. These figures are measured, not estimated — they change as the course material does.
- 279 practice sheets
- 263,823
- 439 handouts
- 408,503
- 439 chat tutor contexts
- 476,741
That is 4 % of a practice sheet. This text fits into a million tokens 23,810 times.
Why a long conversation costs many times more, not twice as much
In a long-running chat, the entire conversation so far is sent again with every round. Consumption therefore does not rise in a straight line — it gets steeper and steeper. Only input is counted here; the output is the same size either way.
- One long chat
- 101,200
- input tokens after all rounds
- A fresh chat per task
- 5,400
- input tokens after all rounds
- Difference
- 18,7 ×
- 95,800 tokens go on re-sending the conversation alone.
After 20 rounds, the long chat has spent 95,800 tokens simply reading itself back — that is 8,3 % of the entire T-AI-L course material.
Why caching only pays off from the second time
A cached prompt costs more than usual the first time and far less after that. Everything here is in multiples of a normal input — the surcharge for writing and the discount for reading are factors, not amounts.
One block = one request. Height = multiple of a normal input.
- Without caching
- 6.00 ×
- inputs in total
- With caching
- 1.75 ×
- inputs in total
- Saved
- 71 %
- less input than without caching
The first call is the expensive one: it costs 1.25 times a normal input. Every further call costs only a tenth. That is why everything recurring belongs at the start of the prompt — past the first change, the cache is worthless.
The three providers charge for different things
What decides day to day is not how high the prices are but what is charged for at all. This table names no amounts — it shows which kinds of cost exist and which lever follows from each.
| What is charged for | Claude | Gemini | ChatGPT | What follows from it |
|---|---|---|---|---|
| Input and output tokens charged separately | yes | yes | yes | Output is a multiple of input everywhere. Short answers are the first lever. |
| Reasoning tokens | counted as output everywhere | Thinking is billed at the most expensive rate. Raise the effort only where it earns its keep. | ||
| Writing to cache costs more than normal input | yes | yes | yes | The first call costs extra. Caching only pays off on repetition — see instrument 3. |
| Reading from cache costs less | yes | yes | yes | Recurring system prompts and documents belong at the start, not at the end. |
| Cache storage time as its own line item | no | yes | no | With Gemini a cache costs even while it just sits there. How long it lasts is a cost question. |
| Price tiers by prompt length | no | yes | yes | For very long prompts this is a structural difference, not a discount detail. |
| Batch processing cheaper | yes | yes | yes | Whatever is not needed right away belongs in the batch. The strongest lever after caching. |
| Faster execution at a premium | yes | – | yes | Speed is a kind of cost. Whoever can wait pays less. |
| Built-in tools charged separately | yes | yes | yes | Web search and code execution appear on the bill next to the tokens. |
| Region or data residency at a surcharge | yes | – | yes | Where the computing happens can change the price. Check it alongside any data protection commitment. |
The three rows where they genuinely differ are storage time, length tiers and the speed premium. Everything else all three have — just under different names.
Token Awareness is part of every package, including the least expensive one. Build your package, or look at a real learning unit.