Video summary
You're Paying Anthropic 20x MORE Than You Need To
Main summary
Key takeaways
Summary
The video explains five ways to reduce token costs when using Claude Code, emphasizing that prompt caching can matter far more than simply asking the model to be concise.
-
Understand prompt caching. Each new message includes the preceding conversation, so input-token usage can grow substantially over a long session. The speaker says Claude Code’s prompt cache makes rereading earlier context much cheaper, but the cache expires after one hour without activity. The video compares cached reads at roughly $1 per million tokens with cache writes at about $20 per million, describing the difference as up to 20×. Changing models or effort settings, switching quick mode, changing MCP servers, plugins or tools, compacting the conversation, or updating Claude Code can also invalidate the cache and make the next request more expensive.
-
Manage a “cold” or expired cache. For a long conversation, the speaker offers three alternatives to simply continuing with a costly fresh cache:
- Clear the conversation and let Claude inspect the project files to regain context.
- Compact the conversation with Claude Code’s built-in feature, which creates a summary for the new conversation.
- Transfer context into a separate Markdown file on disk that Claude can read in a later session.
The speaker recommends not waiting for automatic compaction, particularly in very long sessions, where “context decay” may become an issue.
-
Route tasks to suitable models. Use cheaper or smaller models for tasks that do not require the strongest model. One approach is an advisor/executor setup: a powerful model plans, while a smaller model carries out the work and can consult the larger model if it gets stuck. The video also mentions delegating through tools such as a Codex plugin, using other providers’ models, or using local models where appropriate.
-
Keep Claude Code configuration tidy. The speaker recommends running
/doctorrather than automatically deletingCLAUDE.md. The command can help shorten unnecessary instructions and identify excess context from skills, MCPs, and other configuration. The video notes that a session may already have consumed substantial context before any user messages are sent, and suggests removing unused or overlapping skills. -
Consider token-saving skills and frameworks. The speaker discusses tools such as Ponytail, which aims to reduce the amount of code generated, and Caveman, which is claimed to reduce initial token usage. Shorter instructions such as “be concise” may also reduce output tokens, but the speaker considers these approaches less impactful than prompt-cache management and says results can vary by project.
The video’s central recommendation is to pay close attention to cache behavior: losing a cache during a large session can have a much greater cost impact than small reductions in generated text. Pricing figures and tool results are presented as the speaker’s explanation and examples.
Main Speakers and Sources
- Main speaker: Chase AI narrator.
- Sources discussed or cited: Claude Code documentation, Anthropic pricing information, an advisor-mode blog post, Boris Cherny (identified as Claude Code’s creator), and GitHub repositories and benchmarks for token-saving tools.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.