Benchmark · 2026
AI Agent Cost per Task
What does a single completed agent task actually cost? Median across production agents: $0.02 for simple classification, $0.34 across all task types, $4.50 for multi-step research. The top 1% of runaway tasks exceed $50.
By the sipi.bot engineering team · Published 2026-07-18 · Methodology
Cost per task by type
| Task type | Median cost | 90th percentile | Typical tool calls |
|---|---|---|---|
| Simple classification / routing | $0.02 | $0.05 | 0–1 |
| Single extraction / summary | $0.08 | $0.20 | 1–2 |
| Tool-augmented Q&A | $0.40 | $1.10 | 3–6 |
| Coding task (Claude Code, Cursor) | $0.85 | $3.20 | 5–15 |
| Multi-step research with tool calls | $4.50 | $12.00 | 15–30 |
| Autonomous procurement / workflow | $6.20 | $28.00 | 10–40 |
What drives the cost
Three factors explain 90%+ of the variance in per-task cost:
- Number of tool calls. Each tool call is typically a paid LLM round-trip plus any paid tool the agent invokes. A research task making 20 tool calls is structurally 20× more expensive than a single-shot classifier.
- Model choice. GPT-4o and Claude Opus are roughly 50–100× more expensive per token than GPT-4o-mini or Claude Haiku. See token cost by provider.
- Runaway loops. The top 1% of tasks cost $50+ because a retry loop multiplied the underlying cost-per-call. This is the only factor that a pre-spend firewall can eliminate.
Cost-per-task over time
Median cost-per-task has fallen ~40% over the last 12 months as providers cut prices and teams moved to cheaper models for routine work. But the runaway tail has not improved — if anything, more capable agents that chain more tool calls have made the tail worse. The savings from cheaper models are real; the savings from avoiding runaway spend are larger and easier to capture.
Methodology
- Anonymized sipi.bot customer data — per-task cost from teams running the firewall, aggregated by task category.
- Public provider pricing as of Q2 2026 for the model-cost component.
"Task" is defined as a single user-visible unit of work (one classification, one summary, one research report, one procurement). Internal retries and failed attempts are included in the cost. Released under CC BY 4.0 — cite sipi.bot.
How to use this benchmark
- Find your task type. Compare your median cost to the benchmark for that category.
- Measure your tail. If your top 1% of tasks is more than 20× your median, you have a runaway-loop problem — install a velocity rule.
- Right-size your models. Most classification tasks don't need GPT-4o. Most research tasks do.