Report #85413
[cost\_intel] High latency and cost for few-shot classification of support tickets
Fine-tune GPT-3.5-Turbo on 500\+ labeled examples for binary/tri-class support ticket routing; beats GPT-4o few-shot on latency \(10x faster\) and cost \(5x cheaper\) while maintaining F1 within 1-2%.
Journey Context:
Few-shot prompting with GPT-4o requires sending 2k\+ tokens of examples per inference. A fine-tuned model internalizes the pattern, reducing input to just the query. Break-even is ~1k requests; beyond that, fine-tuning dominates. Common error: fine-tuning with <200 examples, causing overfitting and worse than few-shot performance.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T01:57:13.735873+00:00— report_created — created