Use-case preset
Sentiment analysis cost calculator
Classify reviews or messages by sentiment; batch, high-throughput workload.
High-throughput batch classifier: ingest a review, support ticket, or social post and emit a label (positive/negative/neutral, or a 1–5 star score). Runs offline over millions of records; cost-per-call and throughput are the only metrics that matter.
The 95/5 ratio is near-maximal — the full text of the item dominates the prompt while the output is a single token or short label. A 1k context window is sufficient for the vast majority of reviews and short messages. Latency is batch/best-effort; this workload runs on queues, not in real-time APIs. `cachedPromptPercent` is ~35: the classification schema, label definitions, and few-shot examples repeat across every call. The biggest cost lever is model size — a 1B–3B model fine-tuned on classification can match a 70B zero-shot model at 20–40× lower cost. If you're running >1M items/day, that difference is worth a fine-tuning investment.