1In one sentence
Beyond "send a message, get an answer", the API has a set of features that cut cost, make answers reliable, and let Claude use tools; these are the ones that turn a demo into a product.
2New words
- Prompt caching — if the same long text (your instructions, a big document) goes in every request, Claude remembers it for a while and charges much less for the repeated part.
- Batch — send many requests at once to be answered within hours instead of seconds, at half the price.
- Structured output — you give the shape (a JSON schema, a strict template) and Claude's answer always fits it, so your program can read it without guessing.
- Tool use — you describe functions your program has (look up an order, send an email) and Claude asks to call them with the right inputs; your code runs them.
- Thinking and effort — settings for how hard Claude reasons before answering; more for hard tasks, less for simple ones.
- Context — same as in Chapter 0.5: everything in the request. Big models take up to a million tokens; features like compaction summarize old parts so long jobs keep going.
3Why you would want it
The same question costs ten times less with caching and batching. Structured outputs remove a whole class of bugs. Tool use is how Claude stops answering about your data and starts acting on it. Knowing these six makes the difference between a Claude feature that costs too much and one that pays for itself.
4The map
(as Anthropic groups them, Oct 2026)
| Group | Features | What they give you |
|---|---|---|
| Reasoning | Adaptive and extended thinking, effort, fast mode, task budgets | Right depth per task; speed when you need it |
| Context and cost | Prompt caching, batches, context editing, compaction, message threads (beta: the server keeps the conversation, so you send only new messages), Files API, PDF and image input | Lower bills; long documents; long-running jobs |
| Tools | Custom tools, strict tool inputs, parallel calls, tool search, code execution, web search and fetch, browser and computer use, memory tool | Claude that acts, not just answers |
| Output | Structured outputs, citations, streaming | Answers your program can trust; sources for document questions |
| Operations | Service tiers, spending limits, usage analytics, data residency, tracing | Control in production |
| Connectivity | MCP connector, Agent Skills (with built-in PowerPoint, Excel, Word, PDF skills) | The same skills and connectors as the app |
5Try it now
(developer account)
- Take your script from Chapter 6.6 and put a long instruction (a page of rules) in the system prompt. Send it twice; compare the token counts and cost in the console with caching on.
- Ask for structured output: Return JSON with fields name, date, total. Feed it a receipt text and check the JSON parses.
- Define one tool (for example, get\_weather with a city input) and watch Claude ask to call it.
6What you will see
[Picture pending: console usage showing cached vs uncached cost, and a tool-call response.] The second cached request costs a fraction of the first. The structured answer is valid JSON every time. The tool request comes back as a call with the inputs filled, waiting for your code.
7Watch out
- Features change monthly; the feature overview page in Anthropic's docs is the source of truth. This chapter is a map, not the territory.
- Measure before and after. The cost features only pay when the repeated part is actually repeated.
8Go further
Combine caching + batches + the smallest model that passes your quality check; that is the standard recipe for bulk processing. Chapter 6.8 puts the tools together into an agent.