Claude Opus is brilliant. It's also expensive. At $15/million input tokens and $75/million output tokens, using it for every task is like taking a helicopter to the grocery store.
The Model Tier Strategy
Anthropic offers three tiers with dramatically different pricing:
| Model | Input (per 1M) | Output (per 1M) | Best For |
|---|---|---|---|
| Claude Haiku | $0.25 | $1.25 | Classification, extraction, simple Q&A |
| Claude Sonnet | $3.00 | $15.00 | Analysis, coding, multi-step reasoning |
| Claude Opus | $15.00 | $75.00 | Complex research, long-form writing, nuanced tasks |
Haiku is 60x cheaper than Opus for input tokens. For many agent tasks — intent classification, entity extraction, simple summarization — Haiku performs just as well.
Implementing a Router
The simplest approach: use a cheap model to classify task complexity, then route to the appropriate tier.
# Use Haiku to classify, then route
classification = call_haiku("Classify this task as simple/medium/complex: " + task)
model_map = {
"simple": "claude-3-5-haiku-20241022",
"medium": "claude-sonnet-4-20250514",
"complex": "claude-3-opus-20240229"
}
result = call_model(model_map[classification], task)The classification call costs fractions of a cent and saves dollars on every task that doesn't need Opus.
Measuring the Impact
Track costs per model in AgentBurn before and after implementing routing. Based on published pricing differences, even basic two-tier routing (Haiku + Sonnet) can reduce Anthropic spend substantially — the exact savings depend on your task distribution.
The dashboard's model breakdown chart makes this immediately visible — you'll see your cost distribution shift from a single expensive model to a mix of cheap and expensive, with the total trending down.