Context Window Calculator

The table above converts token limits into familiar units, but the more important conversion is into money and behavior. A 128k window holds roughly a full novel; filling it entirely on a premium model costs real money on every single request — at a $2.50 per million input rate, a maxed-out 128k request is about 32 cents, which at ten thousand requests a month is thousands of dollars. Big windows are capacity, not free storage.

Usable capacity is also smaller than the headline number: the window is shared by the system prompt, the conversation history, your input and the reserved space for the model's answer. An application that lets history grow unchecked will hit the ceiling — or the budget — long before a single document ever would.

Long-context research keeps finding the same pattern: models recall information near the beginning and end of the context far better than material buried in the middle. Stuffing a window with everything you have doesn't just cost more — it can make answers worse, because the relevant passage competes with thousands of irrelevant ones.

Two practical habits follow. Put instructions at the start and the most decision-relevant facts near the end of the prompt. And when your source material is large, retrieval — indexing content and sending only the passages that matter — usually beats window-stuffing on both quality and cost. Treat the window as a ceiling, not a target.

Short-turn chat and classification live comfortably in small windows; long-document Q&A and codebase analysis genuinely benefit from 200k and up; the million-token tier earns its keep when whole repositories or document collections must be reasoned over in one pass. Check what your documents translate to with the pages-per-context guide, and measure your real prompts in the token calculator before assuming you need the biggest window on the menu.

Window size is the first constraint to check and the easiest to get wrong, because the headline numbers are not comparable to each other in practice. As of August 2026, GPT-5.6 Sol, Terra and Luna all carry a 1.05M-token window. Gemini 2.5 Pro and Flash, GLM-5.2, Kimi K3, Nemotron 3 Ultra and Llama 4 Maverick sit at 1M. The Claude family — Opus 4.8, Sonnet 5 and Haiku 4.5 — stops at 200k, roughly a fifth of the others.

Price changes the calculus. Filling a 1M-token window once costs $5.00 on Sol, $0.20 on Luna (down from $1.00 before the 30 July repricing), $1.25 on Gemini 2.5 Pro (though prompts above 200k are billed at the higher $2.50 tier) and $0.20 on Llama 4 Maverick. Filling Claude's 200k window on Opus 4.8 costs $1.00. So the cheap-per-token models are also the ones where a large window is affordable to actually use, while a million-token prompt on a flagship tier is a deliberate, expensive choice rather than a default.

Once you know what fits, price it in the OpenAI or Claude calculator.

A 40-page contract runs about 20,000 tokens. It fits comfortably in every model here, Claude's 200k included, with room for instructions and a long answer. This is the ordinary case, and window size is simply not the constraint.

A 300-page technical manual is around 200,000 tokens. That is the entire Claude window with nothing left for the question or the answer, so in practice you need to retrieve the relevant chapters instead of pasting the book. On a 1M-token model it fits with room to spare — and costs $1.00 on GPT-5.6 Sol just to read it once.

A 20-turn support conversation with a 1,500-token system prompt grows to roughly 15,000–25,000 tokens by the end, because every turn resends the whole history. Cheap per call, but it compounds across thousands of sessions.

A medium codebase, say 50,000 lines, lands near 600,000 tokens — it fits in a 1M window on paper, but retrieval almost always beats pasting it, for both cost and accuracy. The what actually fits guide works through more of these.