Why your Cline bill hit $200 a month, and how to cut it

Open Cline. Type "add a settings page". Walk away for ten minutes. Come back to a working pull request -- and a charge on your provider bill that is much larger than the prompt you typed.

If your monthly Cline spend has reached $200 or more, it is rarely because of one expensive task. It is the way an agent spends tokens.

What changed in this post

The first version gave a percentage breakdown of a "typical" Cline session (with tool output adding up to 74% while the text said 92%), a before-and-after test with exact token counts, and savings figures for each Caveman level. We cannot source any of those numbers, so they are gone. The mechanism below is what matters, and Cline will show you your own numbers.

Where the money goes

A Cline task is a loop: the model asks for a tool, Cline runs it, and the result goes back to the model with everything that came before. Each step is a new API request carrying:

• the system prompt and tool definitions
• your request
• every file Cline has read so far
• every command output it has seen
• every reply the model has written

So the cost of a step grows with the length of the task, and a file read early on is paid for again on every later step until the context is trimmed. That is why a long task costs far more than twice a short one.

The usual heavy items:

• File reads. Large source files, read whole, often more than once.
• Command output. npm install, npm run build and test runs are mostly progress lines and noise.
• Directory listings. Recursive listings that wander into node_modules or build output.
• git diff. A diff across many files can be very long, and agents re-read diffs to check their work.
• Search results. A broad search returns many matches per file.

How to tell: Cline's task header shows the tokens and API cost of the current task. Watch it climb on a long task and you will see that the late steps cost the most.

What does not work

Switching to a cheaper model and changing nothing else. It helps per token, but the loop still resends everything, and a weaker model that needs more steps can end up costing more.

Shorter prompts. Your prompt is a tiny share of each request. The bulk is context Cline gathered.

Clearing the task halfway. You lose the context and pay to rebuild it.

Fix 1: compress tool output (RTK)

kRouter's RTK filter is on by default and rewrites tool results before the request leaves your machine:

It is lossy on purpose, and it only touches content the request marks as a tool result. If your Cline version sends tool output as ordinary message text instead, RTK leaves it alone. kRouter's README estimates 20-40% fewer input tokens for agent work; that is an estimate, not a measurement. The RTK post has the details.

Fix 2: shorter answers (Caveman)

kRouter's Caveman saver adds a system prompt asking the model to answer tersely. It is off by default, has three levels (lite, full and ultra), and the dashboard describes it as about 65% fewer output tokens, up to 87% -- a claim, not a benchmark. Output is usually the smaller part of a Cline bill, so expect a smaller effect than RTK, and shorter explanations along with it.

Both are on the dashboard's Token Saver page.

Fix 3: send routine steps somewhere cheaper

The bigger lever is where the tokens go. Point Cline at kRouter and use a combo that tries free or already-paid access first:

1. kr/claude-sonnet-4.5    # Kiro free tier (50 credits a month)
2. gh/claude-sonnet-4.6    # a Copilot subscription you already pay for
3. glm/glm-5.3             # cheap metered overflow

Cline works through the free and subscription entries first and reaches the metered one only when they are out or failing. How much that saves depends entirely on how much free and subscription capacity you have, which is why we will not promise a number. Kiro and Copilot sessions are licensed for their own clients; kRouter shows a risk notice before you connect them, because a provider can restrict an account it sees used through a proxy.

The setup

npm install -g @sifxprime/krouter
krouter -t

In Cline's settings:

• API Provider: OpenAI Compatible
• Base URL: http://localhost:20128/v1
• API Key: a key from the dashboard's Endpoint page (from the same machine, any value works)
• Model ID: a model such as kr/claude-sonnet-4.5, or your combo's name

If you use the Cline CLI, the dashboard's CLI Tools page can write this configuration for you.

Common questions

Why is Cline so expensive?

Because every step of a task is a new request that resends the whole conversation, including every file read and command output so far. Long tasks get expensive at the end, not the start.

Does a cheaper model fix it?

Partly. It lowers the price per token, but the loop still resends everything, and a model that needs more steps can cost more overall.

Does RTK change what Cline does?

It shortens tool output the model sees: long diffs, file reads and logs are cut and lists are capped. Usually that is noise. If the model needs a part that was cut, it reads it again.

How much will kRouter save me?

It depends on how much of your traffic is tool output and how much free or subscription capacity you can route to. kRouter's README estimates 20-40% fewer input tokens from RTK; routing to access you already pay for can matter more.

Related

• RTK token compression: what it trims
• Cline vs Cursor
• Free AI coding assistant: what is actually free