Optimize
The optimization playbook
Every other hub on this site explains what Claude costs. This one is about the part you control, which is most of it.
Figures last verified 17 September 2026
Nothing here requires a different plan or a cheaper model. These are changes to how a session is built — and because input dominates almost every bill, they are usually worth more than any pricing decision you could make. The mechanics behind that claim are in tokens and context.
The four levers, in order of how much they typically return:
- Session structure. Start a new conversation when the task changes. Carry a summary, not a history. This is the big one.
- Context discipline. Send the function, not the file. Excerpt documents. Put standing rules in project instructions once.
- Model choice. Match the model to the difficulty of the task rather than defaulting to the heaviest.
- Output control. Ask for diffs, not rewrites. Cap length. Stop a wrong answer early instead of letting it finish.
Knowing the levers is not the same as knowing which one you personally are losing money on. That is what the audit is for: your actual transcript, read by a human, with the expensive habit named.
Tools
In this hub
- Prompt caching pricingWrite vs read costs, TTL, minimums, and break-even math.
- What is prompt cachingPlain-language savings range that funnels into the pricing mechanics page.
- 11 ways to cut token usageEleven tactics with before/after token counts — the conversion checklist.
- Reduce Claude Code token usageCoding-loop habits: context, file reads, /clear, subagents, caching.
- Hidden system prompt token costsHow bloated system/project instructions multiply cost every turn — and when to cache vs shorten.
- LLM cost frameworkMeasure → Attribute → Reduce → Monitor: the team-level cost loop.
- Model routing (Haiku vs Opus)When to use Haiku instead of Opus: task-type defaults, quality checks, and what the downgrade saves.
- Tool use costsAgent and tool schema tax: rounds × tokens, why loops multiply, and the levers that cut it.
- Computer use and browser agentsScreenshot-and-action loops bill vision plus text every step — measure steps × tokens and cut the tax.
- Extended thinking costsReasoning tokens billed as output — and how budgets and effort control the spend.
- Document token costsPDF and document ingestion token tax: text vs vision, repeated context, and extract-once fixes.
- Skills and plugins: cost vs convenienceListing vs invocation token cost, and how to trim skills without losing the help.
- Conversation compaction and memoryServer-side summaries keep long threads alive — but the summary pass is billed. When it saves, when it costs.