Claude Code API Pricing: Pay-Per-Token Costs Explained

A token is the fundamental unit of text that language models process. Tokens are not words - they are pieces of words, punctuation, and whitespace that the model's tokenizer splits text into. For English text, one token is roughly 3/4 of a word, meaning 1,000 tokens equals approximately 750 words. Code tends to tokenize less efficiently than prose because of special characters, indentation, and variable names.

To put this in developer-relevant terms: a typical 100-line TypeScript file is approximately 500-800 tokens. A full React component with JSX might be 300-600 tokens. A package.json file is usually 200-400 tokens. When Claude Code reads your codebase, every file it examines consumes input tokens, and every line of code it writes back consumes output tokens.

The distinction between input and output tokens matters because output tokens cost roughly 5x more across all models. This is because generating new tokens requires significantly more computation than processing existing ones. In a typical Claude Code session, you might send 25,000 input tokens (prompts plus file reads plus conversation history) and receive 8,000 output tokens (generated code and explanations). With Sonnet 4.6, that session costs about $0.075 for input and $0.12 for output, totaling roughly $0.20.

The 5x cost difference between input and output tokens reflects the computational difference between reading and writing. Processing input tokens (your code and prompts) is relatively efficient because the model processes them in parallel. Generating output tokens (Claude's code and responses) requires sequential computation where each token depends on all previous tokens, making it much more expensive.

Context accumulation is the biggest hidden cost driver in Claude Code. Every message you send includes the entire conversation history as input tokens. A 3-message conversation might send 5,000 input tokens per message. By message 30, you could be sending 100,000 input tokens per message because the full conversation context is resent each time. This is why long conversations cost exponentially more than starting fresh sessions.

The practical impact: a 5-message session might cost $0.05 total, but extending that same session to 50 messages could cost $3.00 or more, not because each message costs more but because the cumulative context grows. This is why the /compact command and starting fresh sessions are among the most effective cost-saving techniques for API users.

Prompt caching is one of the most powerful cost-saving features available on the Claude API. When enabled, content that remains the same across requests (system prompts, CLAUDE.md files, previously read file contents) is cached and subsequent reads cost only 10% of the normal input token rate. For Sonnet 4.6, cached tokens cost $0.30 per 1M instead of $3.00.

Claude Code takes advantage of prompt caching automatically in many scenarios. Your CLAUDE.md project file, which gets sent with every message, is cached after the first request. File contents that Claude reads remain cached for the duration of your session. The system prompt that configures Claude Code's behavior is also cached.

The savings compound over longer sessions. In a 20-message conversation where your CLAUDE.md is 2,000 tokens and you reference the same 5 files (totaling 10,000 tokens), those 12,000 tokens are cached after the first message. Over 19 subsequent messages, you save approximately $0.60 with Sonnet 4.6 on the cached content alone. For developers with large CLAUDE.md files or projects that repeatedly reference the same modules, prompt caching can reduce input costs by 40-60% over a full working day.

The key question for many developers: when does a subscription become cheaper than the API? Anthropic reports that the average Claude Code API user spends roughly $6 per day. Over 22 working days, that is approximately $132/month, making the Max 5x plan at $100/month a clear saving of $32/month for average-usage developers.

The break-even point for Pro ($20/month) is about $1/day in API costs, which corresponds to very light usage - roughly 2-3 short sessions daily using Sonnet 4.6. If you consistently spend more than $1/day on the API, Pro is cheaper. If you consistently spend more than $4.50/day (roughly $100/month for 22 working days), Max 5x becomes the better deal.

The Claude batch API processes requests within a 24-hour window instead of providing real-time responses. In exchange for this flexibility, every token costs 50% less. Sonnet 4.6 drops from $3/$15 to $1.50/$7.50 per million tokens. Opus 4.6 drops from $5/$25 to $2.50/$12.50.

For Claude Code users, the batch API is ideal for tasks where you do not need immediate results. Common batch-friendly workflows include: running code reviews on pull requests overnight, generating documentation for your codebase, writing test suites for existing code, performing security audits, and analyzing large codebases for refactoring opportunities. If you can submit the work at the end of your day and review results the next morning, you save 50% on every token.

To use the batch API, you submit requests programmatically through the Anthropic API rather than through the interactive Claude Code CLI. This requires some setup but the cost savings are substantial for teams with predictable, non-interactive workloads.

Using Claude Code with your own API key is straightforward. First, create an account at console.anthropic.com and generate an API key. Then set the ANTHROPIC_API_KEY environment variable in your shell configuration. When you run the claude command in your terminal, it will detect the API key and bill usage directly to your API account instead of requiring a subscription.