Architectural Cost & Data Flow Simulator

Turn Cost
$0.0000
Status:   Waiting to start. Click "Turn 1" to begin.
Client
Frontend UI
Cloudflare
Worker Proxy
Search API
Zendesk
Model Provider
Anthropic (Sonnet 4.6)
CACHE ENGINE Empty
REASONING / LLM
New Query (~100t)
3 Zendesk Articles (~1.5kt)
📦 10k KB Material [ephemeral]
Chat History (0t)
Detailed Answer + SE Notes

⏱️ GPU 5-Min TTL

Anthropic enforces a strict Time-To-Live (TTL) of 5 minutes to save GPU VRAM.

• Follow-up < 5m: Cache is read (90% discount), and timer resets.

• Wait > 5m: Cache is completely flushed. You pay the 25% premium to write the 10k tokens again.

📈 The "History Tax"

Your worker uses MAX_HISTORY = 4. This means it sends the last 4 messages (2 Q&A pairs) on every request.

The Math:
• A detailed SE answer is ~1,500 tokens.
• By Turn 3, you are sending ~3,000 tokens of uncached chat history.
• The assistant's long answers hurt your costs more than the user's short questions!