Models with context windows of 200K tokens or more — up to 2M — for whole-codebase and multi-document workloads
| Features | |||||
|---|---|---|---|---|---|
| $0.10 | $0.40 | $0.02 | |||
| $0.40 | $1.60 | $0.10 | |||
| $0.40 | $1.60 | $0.10 | |||
| $15.00 | $60.00 | $7.50 | |||
| $15.00 | $60.00 | $7.50 | |||
| $2.00 | $8.00 | $0.50 | |||
| $2.00 | $8.00 | $0.50 |
Every model on this page accepts at least 200,000 tokens of context — roughly 150,000 words — and the largest stretch much further: Grok 4.1 Fast at 2 million tokens, with Gemini, Claude Sonnet 5, GPT-5.4, DeepSeek V4, and GLM-5.2 at or above the million-token mark. That's enough to fit an entire codebase, a legal document set, or months of chat history into a single prompt.
Advertised size isn't everything: retrieval quality can degrade well before the window is full, and long prompts get expensive fast. Cached input pricing — shown in the list — matters more than the headline price when you re-send large contexts on every request.
Grok 4.1 Fast currently leads with a 2 million token window. Gemini models run just over 1 million, and Claude Sonnet 5, GPT-5.4, DeepSeek V4, GLM-5.2, and Qwen3.7 also offer million-token windows.
Roughly 150,000 English words — about 600 pages. A million-token window fits around 750,000 words: several full-length books, or a mid-sized codebase.
Not uniformly. Most models recall the start and end of a prompt better than the middle, and effective context is often smaller than the advertised maximum. For critical retrieval over huge inputs, test with your own data and consider chunking plus retrieval instead of one giant prompt.
Use cached input pricing: providers charge a fraction of the normal rate for re-sent, unchanged prefixes, which is exactly the shape of chatting over a large document or codebase. Structure prompts so the big static context comes first and only the question changes.