45 lines
1.4 KiB
Markdown
45 lines
1.4 KiB
Markdown
# The 3x Rule of Agent Costs
|
|
|
|
**May 26, 2026**
|
|
|
|
Here is a rule that will save you from budget surprises: **production agent costs 3x your prototype estimate.**
|
|
|
|
## Why the Multiplier Exists
|
|
|
|
| Phase | Multiplier | What Happens |
|
|
|-------|-----------|-------------|
|
|
| Prototype | 1x | Happy path works perfectly |
|
|
| With retries | 1.5x | Failed tool calls retry, edge cases handled |
|
|
| Production | 3x | Monitoring, error handling, observability, security |
|
|
|
|
## Where the Cost Goes
|
|
|
|
Output tokens dominate — about 70% of total cost. The model's reasoning is the expensive part. Input tokens (context) are about 20%. Cached tokens are 10%.
|
|
|
|
Every tool call costs 3-5x more than the call itself:
|
|
- Planning which tool to use
|
|
- Executing the tool
|
|
- Error recovery if it fails
|
|
- Parsing the result
|
|
- Adding the result back to context
|
|
|
|
## Quick Estimation
|
|
|
|
```python
|
|
def estimate_cost(turns, tokens_per_turn, price_per_m):
|
|
base = turns * tokens_per_turn * price_per_m / 1_000_000
|
|
return {
|
|
"prototype": base,
|
|
"with_retries": base * 1.5,
|
|
"production": base * 3.0
|
|
}
|
|
```
|
|
|
|
## Budget Accordingly
|
|
|
|
If your prototype agent costs $0.10 per task, plan for $0.30 in production. At 10,000 tasks per month, that is $3,000 per month, not $1,000. The 3x rule keeps you honest.
|
|
|
|
---
|
|
|
|
*From Module 6 of the [Agentic Engineering Course](/). The full module covers cost optimization, cascade routing, and pass@k evaluation.*
|