# Reduce Claude Code token usage: what actually works By Steven Van · 2026-10-29 Graphify's guide lists Anthropic's token-saving settings first, then says when a code graph helps and cites mixed public test results. Graphify's team has published a guide to cutting Claude Code token usage, and it puts Anthropic's own settings first and the code graph second. According to the post, most of the cost is not the prompt. It is everything [Graphify]() says Claude Code re-sends on every turn: the conversation so far, CLAUDE.md, tool listings and the files read along the way. The [guide]() was written by Syed Fahad, Head of DevRel. ## Where the tokens go The post names four sources of context: re-sent history, CLAUDE.md (loaded at session start and kept for every request), MCP tools, and file reads. Prompt caching means repeated history is re-read at the cached rate, but it still counts. To see the split, run **/usage** for input, output, cache reads and cache writes by model. On Pro, Max, Team and Enterprise plans it also attributes recent usage to skills, subagents, plugins and individual MCP servers. Run **/context** to see what is filling the current window. ## Anthropic's levers first - Use /clear between unrelated tasks, or /compact with instructions to keep what matters. - Keep CLAUDE.md under 200 lines and move workflow-specific instructions into skills, which load only when invoked. - Disable unused MCP servers with /mcp. MCP tool definitions are deferred by default, but connected servers still add a listing to every request. - Match the model and effort to the job: Sonnet for most coding tasks, and /effort to lower thinking on simple work. - Write specific prompts, and push verbose work such as test runs and log parsing to subagents and hooks. - Be deliberate with agent teams, which Anthropic says use about 7× the tokens of a standard session when teammates run in plan mode. ## Where a code graph fits The post says the remaining cost is exploration, which is what Graphify targets. It builds the graph locally with tree-sitter, at zero model tokens, and Claude Code then asks structural questions such as what calls a function or what breaks if it changes. Query answers are capped at 2,000 tokens by default. Graphify says it helps on cross-file questions in large or unfamiliar repos, and does not help on small repos, single-file edits or files you already know. Setup is pip install graphifyy, graphify install and graphify claude install, then /graphify . in the repo. The last install command adds the CLAUDE.md section and hooks that make Claude Code use the graph. ## Mixed public results The post cites tests that vary. One FastAPI run of 11 questions saw files inspected fall 14.6% while input tokens rose 49.5% and cost rose 9.5%. A YouTube run measured about 18% fewer tokens ($5.16 to $4.29) but was slower. An open issue reports Claude Code alone using the same or fewer tokens on common queries. Two token regressions are fixed: an installer instruction that cost about 12 to 25k tokens per question (removed in PR #891), and repeated hook reminders, bounded per session in v0.9.87 (12 October 2026). The one accuracy figure Graphify publishes is key-fact coverage on ERPNext rising from 70.8% to 82.0% across six questions, which is coverage, not tokens. Its advice is to compare /usage on a few real cross-file questions with and without the graph. ![Graphify]() Graphify Turn a codebase into a knowledge graph your AI assistant can trace — answers come back as graph paths with real file:line citations. [View Graphify →]() Sources [Original source]() [Read Reduce Claude Code token usage: what actually works on Creators Toolbox]() --- Canonical source: https://creatorstoolbox.com/blog/graphify-reduce-claude-code-token-usage-what-actually-works