Agent Token Consumption by Task
Token consumption follows the task: short support turns, long coding sessions, huge research contexts. Here's the honest shape — and why you must measure your own.
The shapes
Support turns: short contexts, repeated per ticket — volume-driven.
Coding sessions: long multi-turn contexts, tool outputs — length-driven.
Research tasks: huge contexts (documents, search results) — context-driven.
Voice calls: tokens per turn across the conversation — both.
The honest caveat
Exact token counts vary wildly by model, prompt, and tool design.
The useful benchmark is YOUR audit log, not industry averages.
What to do
Cap per-call spend so one oversized context can't blow the budget.
Trim context, cache stable prefixes, and watch the log.
At a glance
| Task type | Spend shape | Primary control |
|---|---|---|
| Support | Short × volume | Daily ceiling |
| Coding | Long sessions | Velocity limit |
| Research | Big contexts | Per-call cap |
| Voice | Minutes × turns | Category + ceiling |
FAQ
Are these exact numbers?
No — they're shapes. Measure yours with the audit log.
What's the biggest token sink?
Context size on long tasks — cap per-call spend to contain it.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check