← Back to all guides

Claude API Pricing: Input, Output, Cache Tokens

Claude API pricing starts with one formula

If you are comparing Claude API costs, the important thing is not just the model name. Your bill is shaped by input tokens, output tokens, and, when supported, cache tokens. In simple terms, you pay for what you send, what the model generates, and in some cases the reused context that can be cached at a lower rate.

That means the cheapest setup is not always the smallest model. A long prompt, a verbose answer, or repeated context can change the total cost fast.

What each token type means

• Input tokens: the text you send to Claude, including system instructions, user messages, tool schemas, and retrieved context.
• Output tokens: the text Claude generates in response. Longer answers cost more.
• Cache tokens: repeated prompt content that can be stored and reused so you do not pay full price every time for the same static prefix.

If you only remember one thing, remember this: input is what you send, output is what you get back, and cache is reused input that can reduce repeated costs.

How cache tokens actually help

Cache tokens matter most when your app sends the same long instructions again and again. Common examples include a large system prompt, a stable agent policy, tool definitions, or a fixed knowledge base section. Instead of paying full input cost every request, you can reuse that prefix through prompt caching, which often lowers the effective price and speeds up repeated calls.

This is especially useful for agent workflows, support bots, and coding assistants. If your prompt changes only in the last few lines, caching can save a lot more than switching from one premium model to another.

What usually drives the bill up

• Long context windows: more pasted text, docs, or conversation history means more input tokens.
• Verbose answers: if you let the model explain everything, output tokens rise quickly.
• Repeated static prompts: without cache, you pay full input cost every time.
• Large tool payloads: JSON schemas and function definitions can add hidden token usage.

A practical rule is to optimize the part you control first. Shorten instructions, cap response length, and cache anything that does not change from request to request.

How to choose the right Claude setup

Use this decision guide:

• Need the best reasoning? Choose the strongest Claude model available to you, but keep prompts tight.
• Need cheap, high-volume calls? Prefer a smaller model and keep output short.
• Need repeated agent runs? Focus on caching static context before optimizing anything else.
• Need simple integration? Pick a provider that works with your existing SDK and workflow.

That last point is where 59API is attractive. It provides pay-as-you-go access to Claude and GPT models through the base URL https://api.59api.com, and it is compatible with Claude Code, Codex, and any OpenAI SDK. For teams watching spend, it is one of the cheapest relays, uses native official-quality models with no downgrade, and even includes a referral rebate.

Simple checklist before you ship

• Estimate input tokens for your system prompt, user text, and tools.
• Set a response limit so output does not run away.
• Mark reusable text for caching if your workload repeats the same prefix.
• Compare model choices by cost per request, not just by headline price.
• Test with real traffic and inspect token logs after the first batch of requests.
• Use a low-cost relay like 59API if you want to reduce spend without changing your app architecture.

If you are building on Claude and want to keep costs predictable, sign up for a small test on 59API first and compare your real token usage before scaling up.