LLM Context Window Calculator

A context window is the maximum amount of text (measured in tokens) an LLM can process in a single call — it includes both your input (prompt, documents, history) and the model's output. If your input + output exceeds the context window, the model will either refuse, truncate your input, or produce degraded results. Larger context windows let you process longer documents and maintain longer conversations.

A token is roughly 4 characters of English text — about 0.75 words. So 1000 words ≈ 1300 tokens. Code is more token-dense than prose because of punctuation, symbols, and identifiers. Chinese/Japanese characters are often 1 token each. For accurate token counts in production, use the official tokenizer libraries: tiktoken (OpenAI), claude-tokenizer, or the model provider's API.

Remove unnecessary context from your prompt — only include what's relevant. Use system prompts efficiently. Compress chat history (summarize older turns instead of including full text). For RAG systems, retrieve only the most relevant chunks (top-k) rather than large document sections. Use caching (prompt caching in Claude, cached completion in OpenAI) to avoid re-processing repeated context.

Most APIs return an error (400 Bad Request with a 'too many tokens' message). Some models silently truncate the beginning of your input — this is particularly dangerous because the model won't tell you it dropped context. Claude explicitly rejects oversized requests. Always check the token count before sending, especially for agentic systems that accumulate context over multiple turns.

Individual LLM calls are cheap — often fractions of a cent. But production apps make thousands to millions of calls per day. At 10,000 calls/day with a $0.003 average cost, that's $30/day or $900/month. Costs scale linearly with usage, so optimizing token efficiency (shorter prompts, output caching, right-sizing your model) has a compounding effect at scale.

For simple tasks (classification, extraction, short generation): use fast cheap models like GPT-4o mini, Claude Haiku, or Gemini Flash. For complex reasoning and long documents: use Claude Sonnet/Opus, GPT-4o, or Gemini Pro. For very long documents (>100K tokens): only Claude (200K) and Gemini (1M-2M) work. For cost-sensitive high-volume workloads: benchmark multiple models — cheaper models often match quality on well-structured prompts.