Architecture Cost Simulator SE Internal
Simulate
Turn Cost $0.0000
Waiting to start — click Turn 1 to begin the simulation.
Client
Frontend UI
index.html
Cloudflare KV · ARTICLE_CACHE
Article Cache
2hr TTL
Cloudflare KV · SINGULAR_KB
KB Store
No TTL · permanent
Cloudflare
Worker Proxy
singular-proxy
Search API
Zendesk
Help Center
Model Provider
Anthropic · Sonnet 4.6
claude-sonnet-4-6
Prompt Cache
Empty
Reasoning / LLM
Idle
New Query (~100t)
3 Zendesk Articles (~1.5kt)
📦 KB Material (10k tokens)
Chat History (0t)
Answer + SE Notes
KV Hit ⚡
📦 KB from KV
Packets
Query
Zendesk Articles
Article Cache (KV)
KB from KV
Chat History
Answer
Token counts are estimates for illustration
⏱️
5-Minute Prompt Cache TTL
Follow-up within 5m — cache is read at a 90% discount. Timer resets on every hit.
Wait over 5m — cache is completely flushed. You pay the 25% write premium on 10k tokens again.
SINGULAR_KB — stores the full KB, no TTL, permanent. ARTICLE_CACHE — stores Zendesk article bodies for 2 hours. Two separate KV namespaces, two different caching behaviours.
📈
The History Tax
Worker uses MAX_HISTORY = 4 — last 4 messages (2 Q&A pairs) sent on every request.
Cost breakdown (estimates)
KB write (10k tokens) $0.0375
KB read (cache hit) $0.0030
Zendesk articles (~1.5k) $0.0045
History Tax (2 Q&A ~3k) $0.0090
Output (~500 tokens) $0.0075
💡
Key Takeaways
Turn 1 is always expensive — cache write penalty is unavoidable.
Turn 2 is the ideal scenario — KB practically free, only pay for Zendesk + history + output.
Turn 3 is worst case — cache miss + history tax stacked. Long SE answers hurt more than short questions.