Airbnb fine-tuned Gemma4-E2B classifier matches frontier quality at 1% cost
The counterintuitive result is that longer prompts cannot close the quality gap once a task has a latency budget — task-specific fine-tuning beats prompt engineering for enforcement workloads. Teams running high-volume moderation, compliance or routing classifiers should treat sub-billion-parameter fine-tuned models as the default baseline rather than defaulting to APIs. The published break-even threshold (15–30 calls per minute) gives engineering teams a concrete decision rule for build-versus-buy, and the latency and cost methodology is directly reusable.
