You do not need to understand tokens to use an AI coding tool. You do need to understand them to predict the bill, and the bill is the thing that makes people stop using these tools.
Here is the smallest amount of it that is actually useful.
A token is a chunk of text
Roughly three-quarters of a word in English. "Unbelievable" is a few tokens; "the" is one. Code is denser — punctuation and indentation all count.
Rules of thumb that are close enough:
- A page of prose is about 500 tokens.
- A medium source file is 1,000 to 3,000.
- A million tokens is a decent-sized book, or maybe 300 source files.
Prices are quoted per million tokens, which is why they look small. They are small. You just use a great many of them.
You pay for reading and writing, at different rates
Input is everything sent to the model: your message, the files it looks at, the instructions, and everything already said in the conversation.
Output is what it writes back. Output costs several times more per token than input, usually somewhere between three and ten times.
The instinct is therefore to worry about output. That instinct is wrong, and the reason is the next section.
The bill grows faster than the work
This is the whole thing. If you read one section, read this one.
A conversation with an AI is not a series of separate questions. Each time the agent takes a step, it re-sends everything so far — your original request, every file it has read, every change it proposed, every test result it got back.
So step twelve is not a twelfth of the job. Step twelve carries steps one through eleven inside it.
A task that touches six files and takes twenty steps can send those six files twenty times. That is why a session that feels like it is going badly costs so much more than one that goes well: it is not slightly more expensive, it is on a curve.
The practical consequence, and it is not the obvious one:
Shorter tasks are much cheaper than shorter answers.
Cutting the model's verbosity saves a little. Cutting a twenty-step task into two eight-step tasks saves most of the bill — because you halve the steps and shrink what every remaining step is carrying.
The four numbers on a rate card
You will see four prices per model, not one, and now they make sense:
| What it is | |
|---|---|
| Input | New text sent to the model |
| Output | Text the model writes |
| Cached input | Text the model has seen before, re-sent — heavily discounted |
| Cache write | The cost of putting something into that cache |
The cache is the interesting one. Send the same beginning of a conversation twice and the second read is much cheaper. Good.
The catch: a cache write usually costs more than plain input. So a tool that reorders the conversation between steps, or inserts something near the beginning, pays the expensive write over and over and never earns the cheap read. From the outside it behaves identically. On the bill it does not.
That is why Codoro's model catalogue stores all four numbers per model rather than one. A table with a single input price cannot express this, and a table that cannot express it gets used to make a decision that is wrong.
What to actually do
Set a per-run cap. Codoro's default is $2.00 per task, editable, with a $25.00 per-day backstop. Set it low enough that hitting it tells you something — a cap above your worst case has never fired and is protecting nothing. When it fires the run pauses and offers a cheaper model; it does not silently switch, so the decision stays yours.
Ask for the plan before the code. One cheap step that prevents ten expensive ones, because the expensive sessions are nearly always the ones where the model understood the task differently from you.
Give it a narrower task. The single biggest lever, per the curve above.
Use a cheap model for looking around and an expensive one for deciding what to do. Finding files is mechanical; choosing an approach is not.
Use your own API key if you have one. Zero markup — the provider bills you directly and the key stays in your Mac's keychain. Hosted credits carry a 40% markup on cost, which buys you one bill instead of five and nothing else.
What the plans cost
| Tier | Price | Credit included |
|---|---|---|
| Hobby | $0 | None — your own keys |
| Pro | $20/mo | $20 |
| Pro+ | $60/mo | $60 |
| Ultra | $200/mo | $200 |
Notice the credit equals the price. You are not paying a subscription fee and then buying tokens on top — the subscription is prepaid tokens. Hobby exists so bringing your own key costs nothing at all. Top-ups are $28, $70 and $140, never expire, and never auto-charge.
Status
Codoro is in preview for macOS and there is no public download yet — the download page is a waitlist until a signed build ships. The token arithmetic above is true of every AI coding tool, not just this one, so it is worth knowing wherever you end up.