Anthropic's Claude Platform guide argues that cost can fall without accepting a corresponding loss in application performance. This matters now because agent costs accumulate across long contexts, repeated tool loops, retries, and evaluation rather than one visible completion call.
The guide recommends maximizing prompt-cache hit rates and removing prompt anti-patterns when applications move to newer models. It also presents the effort setting as a control for matching reasoning spend to the difficulty of a task. The mechanism is workload-level tuning: stabilize reusable context, reduce conflicting or obsolete instructions, and reserve expensive reasoning for requests that need it.
Teams should measure successful task cost, cache behavior, retries, latency, and quality together before declaring one configuration cheaper. The evidence is Anthropic's platform guidance and reflects its own models and pricing surface rather than an independent cross-provider benchmark. Watch whether the recommendations produce stable savings on representative production traces after model upgrades and prompt changes.