How to cut AI coding agent token costs: 10 practical ways

Why does an AI coding agent use so many tokens?

Because the whole context is sent again on every call. A model doesn't remember the previous step; the agent re-sends everything it needs, every time.

The last one is the sneakiest. Each session starts from zero, so you (or a long rules file) re-supply the same decisions and conventions every time, and pay for them every call.

10 ways to cut AI coding token costs

Roughly in order of how much they save.

1. Stop re-explaining. Keep decisions, preferences and fixes in persistent memory, and inject only the ones relevant to the task, instead of pasting the same background into every session.
2. Start a fresh session for each task. One endless conversation pays for its whole history on every call. A new task rarely needs the last one's transcript.
3. Compact long sessions. When a session must continue, summarise the history into the facts that matter. Most agents have a command for this.
4. Point, don't paste. Name the files and functions that matter rather than pasting whole files or folders.
5. Trim tool output. Cap long test runs and build logs, show failures not passes, and filter noise before it reaches the model.
6. Disable tools you don't use. Every enabled tool's definition is sent on every call. Unused integrations are pure overhead.
7. Use prompt caching. Providers charge less for input that repeats exactly. Keep stable instructions at the start of the prompt and don't edit them mid-session.
8. Right-size the model. Use a smaller, cheaper model for simple subtasks like renames, formatting or summaries, and save the big one for hard reasoning.
9. Keep rules files short. Instruction files are re-read every session. Move rarely needed detail out, and keep only what applies to most tasks.
10. Measure every change. Track tokens and cost per task, so you know which change actually helped and which only felt like it did.

How much can persistent memory save?

A lot, because it replaces the largest repeated cost: history.

On the LoCoMo benchmark, with the same model answering and judging, retrieved memories kept 97% of the whole-conversation accuracy while reading 82% fewer tokens, at about a fifth of the cost. On a far larger history (about 1.27 million tokens per question, too big for any model), memory read 98.5% fewer tokens and still answered 83% correctly. More in long context vs RAG vs memory.

What shouldn't you cut?

Anything the agent actually needs. Starving the context is a false saving.

• The task and its constraints. A vague prompt leads to a wrong answer and a second, longer attempt.
• Decisions that still apply. Forgetting them means the agent re-does or undoes work.
• The relevant code. Guessing at an interface it hasn't seen produces bugs you pay to fix.

The goal is a small context, not an empty one. See context engineering for AI agents for how to decide what earns a place.

How do you measure token usage?

Per task, not per month.

Most coding agents can show usage for a session, and every model provider's dashboard breaks down spend. Pick a few typical tasks, measure them before and after a change, and keep only what moves the number.

Frequently asked questions

Benchmark figures measured by us on LoCoMo and LongMemEval-M, October 2026, with the same model answering and judging within each comparison; costs at API list prices on the day of the run.