DeepSeek V3: Latency, Batching, and Cost Control

When V3 wins

V3 is a strong default for assistants, triage, summarization, and high-volume user chat where you want lower cost per turn than reasoning-first models.

Batching without surprises

Prefer **explicit concurrency limits** in your worker tier over unbounded `Promise.all` fan-out. Tail latency is usually a queueing problem, not a model problem.

Prompt hygiene

Deduplicate static system instructions on the client, and avoid shipping entire documents when a hashed retrieval snippet is enough. Fewer input tokens directly improves gross margin on usage-based billing.