cocodot
← Back to guides
Local card declined, direct access hard? cocodot does both
Claude CodeUpdated 2026-09

Claude Code Costs: Seven Changes That Actually Lower the Bill

Claude Code sends project context on every call, so its cost structure looks nothing like chat. What heavy use really costs, and seven changes ordered by how much they save — the first two alone typically halve the bill.

TL;DR: Claude Code resends project context on every call, so single requests routinely run into the tens of thousands of tokens. A heavy user burns 20 to 50 million tokens a month. **On a frontier model that is roughly $180 to $450 a month; on the mid-tier model, roughly $70 to $180** — the same work, a materially different bill. Seven changes, in order of impact: **(1) default to the mid-tier model** and escalate only when stuck; **(2) exclude irrelevant directories** so it stops reading your build output; **(3) name the files instead of letting it search**; **(4) start fresh sessions** — context compounds; **(5) ask it to make changes rather than explain them**; **(6) split large refactors** into scoped tasks; **(7) look at where the tokens actually went** before optimizing further. The first two are most of the win.

Default to the mid-tier model

This is the single biggest lever and it takes one setting. The mid-tier model handles the large majority of everyday coding — reading code, making scoped edits, writing tests — at roughly a fifth of the frontier model's rate. Escalate deliberately when you hit something genuinely hard, then come back down. Most people who feel Claude Code is expensive have simply never changed the default.

Stop it reading things that do not matter

An ignore file that excludes build output, dependency directories, lock files, and generated assets removes a surprising amount of context from every call. This is pure waste elimination — nothing you want the model to know lives in a build directory. It costs one file and it applies to every call you make afterwards.

Name the files instead of letting it search

When you describe a problem vaguely, the tool explores to find the relevant code, and every step of that exploration is billed. When you name the file and function, it goes straight there. This is the cheapest habit change available and it also makes the output better, because you have already done the part where you know what you want.

Start fresh sessions

Context accumulates across a session, and every call carries all of it. The fiftieth message in a long session can cost many times the first, for the same size of request. When you finish a task, start a new session for the next one. Long sessions feel efficient and are not.

Ask for changes, not essays

Explanation is output, and output is the expensive side of the meter. Telling it to make the change and keep the commentary short is a real saving on a tool used dozens of times a day. Ask for the reasoning when you actually need the reasoning.

Know where your tokens went

Before optimizing further, look at actual usage rather than intuition. People are consistently wrong about which activity dominated their bill — the session that felt expensive often was not, and a routine loop often was. Measure once and you will know which of the six changes above is worth your attention.

Where the savings are

ChangeTypical savingEffort
Default to the mid-tier model50-60%One setting
Exclude irrelevant directories10-30%One file
Name files explicitly10-20%Habit
Start fresh sessionsVaries, grows over timeHabit

FAQ

Does the mid-tier model produce worse code?

For everyday work, the difference is far smaller than the price difference. Use it as the default and escalate on the specific problems where you notice it struggling — that is a better policy than paying frontier rates for reading files.

Is there any way to avoid resending context?

Not in the general case — the tool needs the context to be useful. What you can control is how much of it is worth sending, which is what excluding directories and naming files are for.

About cocodot

cocodot is a payment and AI access service for developers and cross-border teams in mainland China. It provides US-BIN virtual cards issued by a licensed institution — used to pay for overseas subscriptions and ad accounts — and an OpenAI-compatible AI API gateway for calling Claude, GPT and Gemini from within mainland China. Both share one wallet, funded by Alipay and accounted in USD. Card: $9.9 to open, 3% to load, 0% on spend, $1 per active card per month.

Service scope, pricing and limits →
Claude Code Costs: Seven Changes That Actually Lower the Bill · cocodot