Define the session boundary

Cost per session only works when a session has a clear start and end. Treat one terminal conversation or one focused task block as the measurement unit.

• Use task names, not vague timestamps.
• Close the loop after the work ships or stalls.
• Do not mix unrelated projects in one cost bucket.

Separate visible prompts from hidden overhead

AI coding agents can spend tokens on summaries, command handling, and context management. Those tokens matter because they explain why a quiet session may still show burn.

• Look for background-cost notes.
• Track long sessions separately.
• Reset context when the session is drifting.

Compare sessions by outcome

A $5 session that fixes a production issue is different from a $5 loop that only rewrites tests. Cost per session becomes useful when it is paired with outcome labels.

• Bug fix
• Feature build
• Code review
• Exploration

Turn session cost into a budget habit

After a week, you will know which session types are normal and which are outliers. That is enough to set a soft budget without banning useful experimentation.

• Set warning thresholds.
• Review the top three burns weekly.
• Keep one place where the team can compare notes.

Capture a session's cost automatically

A status line script receives JSON on every update. Two fields make a per-session cost log easy: cost.total_cost_usd (estimated session cost, reset to $0 by /clear) and cost.total_duration_ms (wall-clock time, accumulated across resumes). Have your script also print the session name, then write the last value to a file when you finish a task. That gives you one row per task without manual bookkeeping. The status line also exposes context_window.used_percentage, so you can see when context bloat, rather than output, is driving a session's cost.

• Run /rename before /clear so the finished session is easy to find with /resume.
• Separate one task from the next with /clear, not by opening a new terminal tab. That keeps one task per cost bucket.

Why a quiet session can still cost money

Anthropic's cost docs list the hidden drivers. Every request resends the whole conversation, billed at the cached rate when the cache is warm. The first message after a break longer than the cache lifetime misses the cache and reprocesses everything. A /loop or scheduled task, goal check-ins, cross-session messages and active agent teammates can all start turns while you are idle. /compact is itself a large request. Thinking tokens bill as output tokens. On models with a fixed thinking budget, lowering /effort or setting MAX_THINKING_TOKENS reduces them.

• The /usage Prompt cache line (v2.1.251 or later) shows the cache hit share and misses for the main conversation.
• Anthropic estimates agent teams in plan mode use about 7x the tokens of a standard session.

Cap a headless session's spend

For scripted or CI runs in print mode (-p), the --max-budget-usd flag stops the run once estimated API spend reaches the amount you set. Subagent spend counts toward the cap. It is checked against Claude Code's client-side estimate, so it can differ slightly from your bill. Pair the flag with a smaller model for routine jobs, and keep measuring interactive sessions through /usage.