Anthropic Cost Tuning
Overview
Optimize Claude API spend through model routing, prompt caching, the Message Batches API, and real-time cost tracking. The four biggest levers: model selection (4-19x), prompt caching (10x input), batches (2x), and max_tokens discipline.
Pricing Reference (per million tokens)
| Model | Input | Output | Cache Read | Cache Write |
|---|
| Claude Haiku | $0.80 | $4.00 | $0.08 | $1.00 |
| Claude Sonnet | $3.00 | $15.00 | $0.30 | $3.75 |
| Claude Opus | $15.00 | $75.00 | $1.50 | $18.75 |
Message Batches: 50% off all model pricing for async processing.
Cost Calculator
def estimate_cost(
input_tokens: int,
output_tokens: int,
model: str = "claude-sonnet-4-20250514",
cached_input: int = 0,
use_batch: bool = False
) -> float:
pricing = {
"claude-haiku-4-20250514": {"input": 0.80, "output": 4.00, "cache_read": 0.08},
"claude-sonnet-4-20250514": {"input": 3.00, "output": 15.00, "cache_read": 0.30},
"claude-opus-4-20250514": {"input": 15.00, "output": 75.00, "cache_read": 1.50},
}
rates = pricing[model]
uncached_input = input_tokens - cached_input
cost = (
uncached_input * rates["input"] +
cached_input * rates["cache_read"] +
output_tokens * rates["output"]
) / 1_000_000
if use_batch:
cost *= 0.5
return cost
daily = estimate_cost(500, 200, "claude-sonnet-4-20250514") * 10_000
print(f"Daily: ${daily:.2f}")
print(f"Monthly: ${daily * 30:.2f}")
daily_optimized = estimate_cost(500, 200, "claude-haiku-4-20250514", use_batch=True) * 10_000
print(f"Optimized: ${daily_optimized:.2f}/day")
Strategy 1: Model Routing
def route_to_model(task: str, complexity: str) -> str:
"""Route tasks to cheapest adequate model."""
if task in ("classify", "extract", "route", "validate"):
return "claude-haiku-4-20250514"
if complexity in ("low", "medium"):
return "claude-sonnet-4-20250514"
return "claude-opus-4-20250514"
Strategy 2: Prompt Caching
message = client.messages.create(
model="claude-sonnet-4-20250514",
max_tokens=256,
system=[{
"type": "text",
"text": large_reference_document,
"cache_control": {"type": "ephemeral"}
}],
messages=[{"role": "user", "content": user_question}]
)
Strategy 3: Batches for Non-Real-Time
batch = client.messages.batches.create(requests=[...])
Strategy 4: Spend Tracking
import anthropic
from dataclasses import dataclass, field
@dataclass
class SpendTracker:
budget_usd: float = 100.0
spent_usd: float = 0.0
requests: int = 0
def track(self, response):
cost = estimate_cost(
response.usage.input_tokens,
response.usage.output_tokens,
response.model,
getattr(response.usage, "cache_read_input_tokens", 0)
)
self.spent_usd += cost
self.requests += 1
if self.spent_usd > self.budget_usd * 0.8:
print(f"WARNING: 80% budget used (${self.spent_usd:.2f}/${self.budget_usd})")
if self.spent_usd > self.budget_usd:
raise RuntimeError(f"Budget exceeded: ${self.spent_usd:.2f}")
tracker = SpendTracker(budget_usd=50.0)
Cost Reduction Checklist
Resources
Next Steps
For architecture patterns, see anth-reference-architecture.
