Claude Haiku 4.5: Hız ve Düşük Maliyet Rehberi

1result = client.messages.create(
2    model="claude-haiku-4-5",
3    max_tokens=50,
4    messages=[{
5        "role": "user",
6        "content": f"Bu destek talebini kategorize et: {user_message}\n\nKategoriler: billing, technical, feature_request, complaint, other\n\nSadece kategori adını dön."
7    }]
8)
9# Accuracy: %95, Latency: 300ms, Cost: $0.0008

1result = client.messages.create(
2    model="claude-haiku-4-5",
3    max_tokens=200,
4    messages=[{
5        "role": "user",
6        "content": f"Bu email'den JSON çıkar: sender_name, email, phone, request_type, urgency\n\nEmail:\n{email_text}"
7    }]
8)
9# JSON output %99 valid

1result = client.messages.create(
2    model="claude-haiku-4-5",
3    max_tokens=20,
4    messages=[{
5        "role": "user",
6        "content": f"Bu yorum uygun mu? 'yes' veya 'no' ile cevap ver. Sebep gerekmiyor.\n\nYorum: {comment}"
7    }]
8)
9# Binary classification %97+

1def route_task(user_query: str) -> str:
2    result = client.messages.create(
3        model="claude-haiku-4-5",
4        max_tokens=30,
5        messages=[{
6            "role": "user",
7            "content": f"Bu sorguyu hangi modele göndereceğim? 'haiku', 'sonnet', 'opus' — sadece model adı dön.\n\nSorgu: {user_query}"
8        }]
9    )
10    return result.content[0].text.strip()
11 
12# Haiku routes diğerlerine — $0.001/request

1# 500-word article → 2-sentence summary
2# Quality %94, Cost $0.002

1class ModelRouter:
2    def __init__(self):
3        self.client = Anthropic()
4 
5    async def route(self, query: str) -> str:
6        """Haiku'yu classifier olarak kullan."""
7        decision = await self.client.messages.create(
8            model="claude-haiku-4-5",
9            max_tokens=50,
10            messages=[{
11                "role": "user",
12                "content": self._routing_prompt(query)
13            }]
14        )
15        return json.loads(decision.content[0].text)
16 
17    def _routing_prompt(self, query: str) -> str:
18        return f"""Classify this query complexity:
19- simple: extraction, classification, short QA (→ haiku)
20- medium: coding, content, multi-turn (→ sonnet)
21- complex: multi-step reasoning, large codebase (→ opus)
22 
23Query: {query}
24 
25Return JSON: {{"tier": "simple|medium|complex", "confidence": 0.0-1.0}}
26"""
27 
28router = ModelRouter()
29decision = await router.route(user_input)  # $0.001
30model_id = {"simple": "claude-haiku-4-5", "medium": "claude-sonnet-4-6", "complex": "claude-opus-4-7"}[decision["tier"]]

1response = client.messages.create(
2    model="claude-haiku-4-5",
3    extra_headers={"anthropic-beta": "prompt-caching-2024-07-31"},
4    max_tokens=200,
5    messages=[{
6        "role": "user",
7        "content": [
8            {
9                "type": "text",
10                "text": SYSTEM_INSTRUCTIONS,  # 2k token, cached
11                "cache_control": {"type": "ephemeral"}
12            },
13            {
14                "type": "text",
15                "text": user_query  # variable
16            }
17        ]
18    }]
19)

1import asyncio
2from anthropic import AsyncAnthropic
3 
4client = AsyncAnthropic()
5 
6async def classify_stream(queries: list[str]):
7    tasks = [
8        client.messages.create(
9            model="claude-haiku-4-5",
10            max_tokens=20,
11            messages=[{"role": "user", "content": f"Category? {q}"}]
12        )
13        for q in queries
14    ]
15    return await asyncio.gather(*tasks)
16 
17# 100 parallel classifications in ~500ms
18results = asyncio.run(classify_stream(user_queries))

1batch = client.batches.create(
2    requests=[
3        {
4            "custom_id": f"req-{i}",
5            "params": {
6                "model": "claude-haiku-4-5",
7                "max_tokens": 50,
8                "messages": [{"role": "user", "content": msg}]
9            }
10        }
11        for i, msg in enumerate(messages)
12    ]
13)
14 
15# 24h sonra result'ı al
16results = client.batches.results(batch.id)
17# $0.40/1M input, $2/1M output

1# 1. Haiku: Task classifier (cheap)
2task_type = await haiku_classify(user_input)  # $0.001
3 
4# 2. Task-specific worker:
5if task_type == "faq":
6    answer = await haiku_answer(user_input, faq_context)  # $0.005
7elif task_type == "coding":
8    answer = await sonnet_answer(user_input)  # $0.02
9elif task_type == "complex":
10    answer = await opus_answer(user_input)  # $0.15
11 
12# 3. Haiku: Quality check (cheap verification)
13quality = await haiku_verify(user_input, answer)  # $0.001
14 
15# Total avg cost: ~$0.008 (Haiku heavy, Opus rare)