Token Reduction Techniques

from anthropic import Anthropic

client = Anthropic()

SYSTEM_PROMPT = """[5,000-token product doc + detailed rules...]"""

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": SYSTEM_PROMPT,
            "cache_control": {"type": "ephemeral"},
        }
    ],
    messages=[{"role": "user", "content": "Customer question..."}],
)
# Subsequent calls within TTL hit the cache at a fraction of cost.

# Wrong: free text → long, hard to parse
prompt = "Tell me the sentiment of this review."
# Output: "Sure, the review is quite positive because..."  → 80 tokens

# Right: structured
import json

response = client.messages.create(
    model="claude-haiku-4-5-20251001",
    system='Reply with JSON only. Schema: {"sentiment": "pos"|"neg"|"neu"}',
    messages=[{"role": "user", "content": review}],
    max_tokens=20,  # hard cap
)
result = json.loads(response.content[0].text)
# Output: {"sentiment": "pos"}  → 8 tokens

def smart_route(question, context):
    # Easy, short → Haiku ($0.80/$4 per 1M tokens)
    if len(question) < 200 and not needs_reasoning(question):
        return call_model("claude-haiku-4-5-20251001", question, context)

    # Medium → Sonnet ($3/$15 per 1M)
    return call_model("claude-sonnet-4-6", question, context)

    # Only complex reasoning → Opus ($15/$75 per 1M)

if len(messages) > 10:
    summary = client.messages.create(
        model="claude-haiku-4-5-20251001",
        system="Summarize as bullets. Preserve decisions, names, numbers.",
        messages=messages[:8] + [{"role": "user", "content": "Summarize."}],
    ).content[0].text

    messages = [{"role": "user", "content": f"<summary>\n{summary}\n</summary>"}] + messages[8:]

# Pull top-20, rerank to top-3
candidates = vector_db.search(query, top_k=20)
ranked = reranker.rank(query, candidates)
top_chunks = ranked[:3]

# 3 chunks × 500 tokens = 1500 tokens
# vs naive top-10: 5000 tokens
# 70% savings, often better answers

# Before (245 tokens):
You are a customer service assistant for our e-commerce store. Please
make sure to be polite and professional at all times. When responding
to customers, you should always reply in Turkish unless they ask
otherwise. Your responses should be concise but helpful, and you
should never offer discounts that are not officially approved by our
management team. If a customer asks about something you don't know,
politely inform them and offer to escalate to a human agent.

# After (89 tokens):
ROLE: e-commerce support assistant.
LANGUAGE: Turkish.
TONE: concise, professional.
RULE: never offer unapproved discounts. Escalate unknowns to humans.

RESPONSE STYLE:
Cut filler. Drop articles ("the", "a"). Drop pleasantries.
Use fragments where clear. Code unchanged. Keep technical
substance. One word when one word enough.

def select_tools(user_message):
    intent = classify_intent(user_message)

    tool_groups = {
        "files": [read_file, write_file, list_files],
        "github": [create_pr, list_issues, comment_pr],
        "data": [query_db, run_analytics],
    }

    return tool_groups[intent]

response = client.messages.create(
    model="claude-sonnet-4-6",
    tools=select_tools(user_message),  # 3 tools instead of 30
    messages=[...],
)

batch = client.messages.batches.create(
    requests=[
        {
            "custom_id": f"req-{i}",
            "params": {
                "model": "claude-sonnet-4-6",
                "max_tokens": 256,
                "messages": [{"role": "user", "content": doc}],
            },
        }
        for i, doc in enumerate(documents)
    ]
)

# Results in hours, half off