GuideSave tokens
How to save tokens in Claude Code and Codex.
Most tokens go to the context an agent rereads on every turn: the conversation so far, your instructions, the files it opened and its tools. Saving tokens mostly means keeping that context small and keeping its start the same, so the provider can serve it from cache at a tenth of the price.
AKeep the context small
Start fresh between tasks
Use /clear in Claude Code or /new in Codex when you move to something unrelated. An old conversation is reread on every turn.
Compact long sessions
/compact in either agent replaces the conversation with a short summary, so later turns carry less.
See what fills it
/context in Claude Code shows how much of the window goes to the conversation, instructions, tools and files.
Point at the right files
Name the files or folder to look at. Asking an agent to explore the whole project costs far more than telling it where to start.
Paste the part that matters
Give the ten lines of an error, not the whole log.
BSpend less per token
Use a smaller model for small work
Renames, summaries and quick questions do not need the largest model. /model switches in both agents.
Keep instructions short
CLAUDE.md and AGENTS.md load into every session. Every line there is paid for again and again.
Turn off tools you do not use
Each connected MCP server adds its tool descriptions to the context, even when no tool is called.
Keep the start of the prompt stable
Providers cache the beginning of a prompt. Editing your instructions or tools in the middle of a session breaks that cache and the next turn pays full price.
Hand simple parts to free models
A local model or a free model does small parts of a task without touching your plan.
CWhat Lunchbox does for you
Built for prompt caching
Every room keeps the conversation byte for byte the same between turns and puts what changes, like shared memory and the task, at the end. Providers read the rest from cache, at up to 90% less.
Briefings instead of rereading
A crewmate starts with a short briefing: the goal, what the room learned and its own task. Only the lead reads the whole log.
Free and local models in the crew
OpenCode adds a dozen free models in one click, and local models run on your own computer.
See the saving
The Savings page shows what caching saved, per agent and per room.
DQuestions
How do I reduce token usage in Claude Code?
Clear or compact between tasks, point it at specific files, keep CLAUDE.md short, turn off MCP servers you do not use, use a smaller model for small work, and keep the start of the prompt stable so it is read from cache.
Does prompt caching save tokens?
It saves money and plan usage on input tokens. Cached tokens still count, but Claude and OpenAI bill them at about a tenth of the normal input price.
How does Lunchbox save up to 90%?
It keeps each room's conversation identical between turns so providers serve it from cache, where those tokens cost about a tenth. The saving applies to cached input tokens; new input and output cost the same.