OpenAI Automatic Cached Input: "Sihirli" Otomatik Cache'i Maksimize Etmek

OpenAI Cached Input Fiyatları (Mayıs 2026)#

GPT-5 dramatik indirim#

Cache Hit Tetikleme Koşulları#

Koşul 1 — Min 1024 token prefix#

Koşul 2 — Birebir aynı prefix#

Koşul 3 — Tools list aynı sırada#

Koşul 4 — 5-10 dakikalık TTL#

Koşul 5 — Aynı model versiyonu#

Cache Hit Tespit Etme#

response = openai.chat.completions.create(
    model="gpt-5-2026-02-01",
    messages=[
        {"role": "system", "content": LARGE_SYSTEM_PROMPT},  # 2K+ token
        {"role": "user", "content": user_query},
    ],
)

usage = response.usage
total_input = usage.prompt_tokens
cached_input = usage.prompt_tokens_details.cached_tokens or 0
cache_hit_ratio = cached_input / total_input if total_input > 0 else 0

print(f"Cache hit: {cached_input}/{total_input} = {cache_hit_ratio:.1%}")

Tipik observation#

Cache Hit'i Maksimize Stratejileri#

Strateji 1 — Deterministic Prompt Build#

"You are a customer support assistant.\n"
"You are a customer support assistant. "  ← trailing space
"You are a customer support assistant.\r\n"  ← CRLF

Strateji 2 — System Prompt'u Stabil Tut#

SYSTEM_PROMPT_V12 = """..."""  # bu hash deterministic
# Yeni version eklerken eskini bir hafta paralel tut, gradual rollout

Strateji 3 — Tools Listesini Alfabetik Sırala#

tools = sorted(get_user_tools(user), key=lambda t: t["function"]["name"])

response = openai.chat.completions.create(
    model="gpt-5",
    messages=[...],
    tools=tools,  # ← her zaman aynı sırada
)

Strateji 4 — Statik Bilgiyi Başa, Dinamik Sona#

# YANLIŞ — kullanıcı sorusu ortada
messages = [
    {"role": "system", "content": INTRO},
    {"role": "user", "content": user_query},   # ← dinamik, breaks cache
    {"role": "assistant", "content": EXAMPLE_RESPONSE},
    {"role": "user", "content": ANOTHER_STATIC_PART},
]

# DOĞRU — dinamik kısım sona
messages = [
    {"role": "system", "content": INTRO + STATIC_CONTEXT + EXAMPLES},  # 4K cacheable
    {"role": "user", "content": user_query},  # dinamik, sona
]

Anthropic vs OpenAI Cache — Detaylı Karşılaştırma#

Hangi durumda hangisi daha iyi?#

Pratik Örnek — Aynı Pipeline, İki Sağlayıcı#

def run_with_caching(provider: str, query: str):
    if provider == "anthropic":
        return anthropic.Anthropic().messages.create(
            model="claude-sonnet-4-6",
            system=[
                {"type": "text", "text": SYSTEM_PROMPT,
                 "cache_control": {"type": "ephemeral"}},
                {"type": "text", "text": FAQ + CATALOG,
                 "cache_control": {"type": "ephemeral"}},
            ],
            messages=[{"role": "user", "content": query}],
            max_tokens=500,
        )
    elif provider == "openai":
        return openai.chat.completions.create(
            model="gpt-5-2026-02-01",  # pinned
            messages=[
                {"role": "system", "content": SYSTEM_PROMPT + "\n\n" + FAQ + CATALOG},
                {"role": "user", "content": query},
            ],
            max_tokens=500,
        )