Programmers Worldwide Are Wasting Money on Anthropic! The Company Can't Stand It Anymore
Anthropic recently published guidelines to help developers using Claude Code reduce unnecessary token costs. The key recommendations include: 1) Clear (/clear) conversations after completing a task to avoid carrying irrelevant file reads and command outputs into the next task. 2) Set the model and reasoning effort level at the start of a session, as switching mid-session invalidates the prompt cache, requiring a full-price recalculation of the entire dialog history. 3) Attach files using @ references instead of typing paths manually to avoid extra tool calls and searches that bloat the context. 4) Add quiet flags to verbose commands (e.g., in CLAUDE.md) to minimize lengthy output in the dialog history. 5) Use /compact while the session cache is still warm (before breaks) to compress the dialog at one-tenth the cost. 6) Offload large-output tasks to a sub-agent, which runs in an isolated context and only returns conclusions, preventing intermediate outputs from polluting the main dialog.
The article explains token pricing: input tokens (prefill) are processed in parallel, while output tokens (decode) are generated serially, making output tokens five times more expensive. Caching is crucial for savings—if a request's prefix (system prompt, CLAUDE.md, dialog history) matches the previous one byte-for-byte, reading it costs only 10% of the standard input price. However, cache invalidation occurs when changing models, effort levels, fast mode, compressing dialogs, after cache expiration, or when resuming old sessions. Dialog history also grows quadratically (O(n²)) as file contents and command outputs accumulate, increasing costs per round. Proactive context management—like isolating noisy tasks, using /rewind to trim unproductive turns, and task-based session clearing—is becoming an essential skill for cost-effective AI-assisted development.
marsbitHace 3 hora(s)
