100 emails: 10K tokens - A long article: 5K tokens - A novel: 150K tokens - A whole codebase (small project): 50K tokens

Multi-novel corpus - Year of customer support tickets - Entire law database For these: use RAG.

Bigger context = more relevant info available at inference time. But: - Cost: doubles when you double input - Latency: longer input = slower response - Attention dilution: model may miss details in long context

Lost-in-the-Middle Problem

LLMs attend most to the beginning and end of context. Middle gets lost. Put the most important info early or late.

When to Use Big Context

Multi-file code refactoring - Document analysis (one paper) - Long conversations (chat history)

When to Use RAG Instead

Knowledge base (FAQ, docs) - User-specific data (their history) - Anything dynamic

Tools like LangChain offer summarization to fit more into context. Useful when context is borderline.

Claude Opus at $15/M input: 100K tokens = $1.50. Don't put everything in context — be selective.

For dynamic data: read [what is RAG](https://vibecodingturkey.com/en/rehberler/what-is-rag).

What Is a Context Window 2026 — Guide

Q: Sizes 2026

Gemini 1.5 Pro: 1,000,000 tokens (750K words) - Claude Opus 4.7: 200,000 tokens (150K words) - GPT-5: 128,000 tokens (95K words) - Llama 3.1 70B: 128,000 tokens - Mistral Large 2: 128,000 tokens

What Is a Context Window 2026

The context window is the maximum number of tokens input + output combined an LLM can process at once. Sizes 2026 - Gemini 1.5 Pro: 1,000,000 tokens 750K words - Claude Opus 4.7: 200,000 tokens 150K words - GPT-5: 128,000 tokens 95K words - Llama 3.1 70B: 128,000 tokens - Mistral Large 2: 128,000 tokens What Fits - 100 emails: 10K tokens - A long article: 5K tokens - A novel: 150K tokens - A whol…

All guides

Related guides

What Is Chain-of-Thought 2026
What Is an AI Embedding 2026
What Is Few-Shot Learning 2026
What Is AI Fine-Tuning 2026
What Is Hallucination in AI 2026