A financial services company runs a real-time customer-facing chat assistant on Amazon Bedrock using a large Anthropic Claude model. Users complain that first-token and total response times are too slow during interactive sessions. The team has already enabled response streaming, but perceived latency for the initial reasoning steps remains high. They want to reduce inference latency for this specific interactive workload without changing the model family or degrading output quality, and they are willing to accept a higher per-token price for the faster path. Which approach best meets these requirements?