# Token limits

Source: https://docs.coderhelm.com/concepts/token-limits

Monthly caps on new tokens, cache reads and cache writes, so a busy month never surprises you.

Your team pays Anthropic directly for the tokens CoderHelm uses. **Token limit** (under Account) sets monthly caps so usage can't run away.

There are three limits, because the three kinds of token cost very different amounts:

| Limit | What it counts |
| --- | --- |
| **New tokens (input + output)** | Tokens sent to the model for the first time, plus everything the model writes. |
| **Cache read tokens** | Input the model reads back from its prompt cache. Much cheaper than new input. |
| **Cache write tokens** | Input written into the prompt cache so later calls can read it. |

A run re-reads the same code and instructions many times, so most of its input is cache reads. Splitting the limits lets you set a tight cap on the expensive kind without stopping runs over cheap cache reads.

## How limits work [#how-limits-work]

- Leave a limit blank (or 0) for no limit.
- Each limit is checked on its own. The first one reached pauses new work until the next month starts (UTC).
- CoderHelm sends a warning email when a limit reaches 80%, and again at 100%.
- The status at the top of the page reads **Under the limits**, **Near a limit** or **Limit reached**.

When a limit stops a ticket, CoderHelm comments on it, for example: "You've reached this month's **cache write** limit (… of …). Raise or clear that limit in Settings → Token Limit to continue."

Only admins can change limits. Teams that set a single limit before the split keep it until they save the new fields.

## Example [#example]

A team that wants to spend at most about the price of 20 million new tokens a month, without counting cheap re-reads, might set:

| Limit | Value |
| --- | --- |
| New tokens | 20,000,000 |
| Cache read tokens | blank (no limit) |
| Cache write tokens | 10,000,000 |

See [Usage and audit](/concepts/usage-and-audit) for where each token went.