TermdockStart with Billing and Usage
Claude Code cost depends on authentication, plan, model, and task. An earlier version claimed five tested methods cut costs in half, but provided no verifiable measurement. Treat the practices below as things to test yourself, not a guaranteed percentage saving.
According to Anthropic's cost documentation, /usage displays session usage and an estimated dollar figure. API users should check Console usage and billing; Pro/Max subscribers should inspect plan usage rather than treating the session estimate as an extra charge. Record a representative period before comparing changes.
Method 1: Test Model and Effort Against the Task
Check the current model and supported effort settings. The model configuration guide describes /model, /effort, and their limits. Compare settings on a set of low-risk tasks using the same acceptance criteria, recording tests, retries, and usage. Lower effort does not guarantee task completion or a lower bill.
Method 2: Make Project Instructions Specific and Testable
Put the architecture, constraints, and test commands you actually need in CLAUDE.md. Avoid pasting whole documents, credentials, or every file the agent can already access. After editing, run the same tasks and check whether it found the right files, respected constraints, and ran the right tests. See the CLAUDE.md writing guide.
Method 3: Use Another Tool When It Helps
Exploration or read-only review can go to another CLI you can access, but that tool may add subscription or API charges and requires a handoff. Google's transition announcement conflicts with Gemini CLI's authentication and quota docs on personal-account access; do not assume the old 1,000 free daily requests.
In a Termdock project workspace, have Claude Code implement and report the requirement, diff location, and tests. If useful, pass these to another read-only session. Follow the agent handoff walkthrough and record actual usage on both sides. Agents do not automatically share conversations, and two agents are not guaranteed to cost less than one.
Method 4: Give Subagents Bounded Tasks
Delegate only when an independent check or parallel investigation actually helps. Specify scope, output format, and prohibited actions. The coordinator verifies sources and results before deciding what follows. Extra agents consume resources; for a small task a manual check may be simpler.
Method 5: Manage Context Without Losing Essentials
When the task changes or the conversation grows, consult the Claude Code interactive-mode docs for current context controls. Retain the requirement, file locations, decisions, and open items so compaction or a new session does not lose them; then verify the agent still knows the acceptance criteria.
How to Tell Whether It Helped
Compare similar tasks, actual bills, usage, outcome quality, and manual correction time. A fixed subscription does not refund itself because you use it less once; include the other model's API bill in the total. If the records are insufficient, the conclusion is "not measured," not an automatic 50% saving.