A team is piloting a supported Bedrock model with low, unpredictable traffic and no firm capacity commitment. Which approach is the most reasonable starting point before considering supported provisioned-capacity options?

1. Purchase the largest supported provisioned allocation immediately because it is always cheaper at any utilization.
2. Assume all models and Regions offer identical throughput options and commitments.
3. Select the model with the highest token price because price alone guarantees the best application results.
4. Use an available on-demand inference option, measure usage and latency, and evaluate provisioned capacity if predictable sustained demand justifies it.