How to cut your AI coding token costs

A practical guide for developers using Claude Code, Cursor and similar agents. These are the levers that actually move your bill.

1. Keep context small and deliberate

Tokens are charged on everything the model reads, not just what it writes. The biggest hidden cost is context: large files, whole-repo reads, and long conversations that get re-sent on every turn. Start fresh sessions for unrelated tasks instead of letting one conversation grow to tens of thousands of tokens. Point the agent at specific files rather than asking it to "look around".

2. Use caching where the tool supports it

Modern models support prompt caching: stable, repeated context (system prompts, large reference files) is billed at a steep discount on subsequent calls. Claude Code reports cache-read and cache-write tokens separately — if your cache-read numbers are high, caching is working for you. Structure prompts so the unchanging part comes first and changes go at the end.

3. Match the model to the task

Reach for the largest model only when you need it. Routine edits, renames and boilerplate are well within the reach of smaller, cheaper models that cost a fraction per token. Reserve the flagship model for genuinely hard reasoning. Switching model tiers per task is often the single biggest saving.

4. Watch retries and tool loops

Agentic tools can get stuck in loops — re-reading files, retrying a failing command, re-planning. Each loop is billed. If you see a session's token count climbing fast with little progress, stop it and re-scope. adonterminal's per-session view makes these runaway sessions easy to spot.

5. Trim what you paste

Pasting an entire stack trace, log file, or generated bundle into a prompt is expensive and rarely necessary. Paste the relevant lines. Summarise long inputs before sending them. The model usually needs the signal, not the noise.

6. Review your spend weekly

What gets measured gets managed. A five-minute weekly look at which tools, models and days cost the most will surface patterns no single session reveals — a particular project that's pricey, a model you over-use, a time of day you burn most. That habit, more than any single trick, is what keeps the bill down.