Claude Code Reframes Cost Around Finished Tasks

The key insight is that agent efficiency—not headline token pricing—is becoming the real competitive moat.

• –Fewer turns reduce repeated context processing and lower the chance of costly retries.
• –Prompt caching can cut input costs sharply, but pauses, model switches, MCP changes, compaction, or prompt edits can trigger expensive cache rewrites.
• –Anthropic’s 40% Opus 5.5 savings estimate is workload-dependent, so developers should benchmark their own tasks with `/usage`.
• –Raising effort can be cheaper than switching models when it prevents a failed attempt or multi-turn detour.
• –Claude Code users should optimize for cost per successful task, not cost per request.