What Is a Context Window?
Short answer: the token budget one request shares between everything you send and everything the model writes back. Current models range from 200,000 tokens (Claude Haiku 4.5) to 1,050,000 (GPT-5.6). Filling one completely costs anywhere from $0.20 to $8.40 — a 42x spread driven by window size as much as by price per token. Below: how the budget actually works, what each window holds, and why 100% utilization is usually the wrong target.
One budget, shared by input and output
The most common misconception is that the context window is a limit on the prompt. It is a limit on the whole exchange.
Context windows and output caps, side by side
All eight tracked models. Words and pages assume ordinary English prose at 0.75 words per token and 500 words per page.
| Model | Context window | Max output | Words it holds | 500-word pages |
|---|---|---|---|---|
GPT-5.6 Sol OpenAI · flagship (promo) | 1,050,000 | 128,000 | 787,500 | 1,575 |
GPT-5.6 Terra OpenAI · mid | 1,050,000 | 128,000 | 787,500 | 1,575 |
GPT-5.6 Luna OpenAI · budget | 1,050,000 | 128,000 | 787,500 | 1,575 |
Claude Opus 5 Anthropic · flagship | 1,000,000 | 128,000 | 750,000 | 1,500 |
Claude Sonnet 5 Anthropic · mid | 1,000,000 | 128,000 | 750,000 | 1,500 |
Claude Haiku 4.5 Anthropic · budget · smallest | 200,000 | 64,000 | 150,000 | 300 |
Gemini 3.6 Flash Google · mid (promo rate) | 1,048,576 | 65,536 | 786,432 | 1,573 |
Gemini 3.5 Flash-Lite Google · budget | 1,048,576 | 65,536 | 786,432 | 1,573 |
Context windows are provider-published limits; word and page figures are planning estimates at 0.75 words per token. The gap between the two OpenAI and Gemini windows is 1,424 tokens — 0.14% — while the real structural difference is on the output side, where GPT-5.6 and Claude Opus/Sonnet 5 generate roughly twice what either Gemini model can return in one response.
What it costs to fill one, end to end
Each row prices a single request that fills the window entirely with input, at published per-1M rates. This is the maximum possible cost of one call to that model.
| Model | Tokens to fill | Effective input rate / 1M | Cost to fill | Cost per 1M tokens |
|---|---|---|---|---|
GPT-5.6 Sol OpenAI · long-context rate | 1,050,000 | $8.00 | $8.40 | $8.00 |
GPT-5.6 Terra OpenAI · long-context rate | 1,050,000 | $4.00 | $4.20 | $4.00 |
GPT-5.6 Luna OpenAI · long-context rate | 1,050,000 | $0.40 | $0.42 | $0.40 |
Claude Opus 5 Anthropic · flat rate | 1,000,000 | $5.00 | $5.00 | $5.00 |
Claude Sonnet 5 Anthropic · flat rate | 1,000,000 | $2.00 | $2.00 | $2.00 |
Claude Haiku 4.5Lowest Anthropic · smallest window | 200,000 | $1.00 | $0.20 | $1.00 |
Gemini 3.6 Flash Google · flat rate (promo) | 1,048,576 | $0.75 | $0.7864 | $0.75 |
Gemini 3.5 Flash-Lite Google · flat rate | 1,048,576 | $0.30 | $0.3146 | $0.30 |
Computed from published per-1M input rates (OpenAI verified 2026-08-23, Anthropic and Google 2026-08-09 to 2026-08-16). OpenAI rows use long-context rates because filling a 1,050,000-token window exceeds the 272,000-token threshold where input doubles and output rises 1.5x. Gemini 3.6 Flash's promotional rate runs through Dec 31, 2026. Caching, batch and taxes excluded.
Price any fill level on every model
Most production prompts use a fraction of the window. Set the fill percentage and see what that costs — including where OpenAI's long-context surcharge switches on.
| Model | Context window | Tokens at this fill | Effective input rate / 1M | Cost to fill | Words | 500-word pages |
|---|
On OpenAI, the last two-thirds of the window costs double.
Every tracked provider except OpenAI charges one flat input rate across the whole window. GPT-5.6 models do not: above 272,000 input tokens, input doubles and output rises 1.5x. Fill a 1,050,000-token window and you pay $8.00 per 1M on Sol instead of the promotional $4.00 — the last 778,000 tokens are billed at twice the rate of the first 272,000.
The practical consequence is counter-intuitive: splitting a long document into sub-272K chunks is cheaper than sending it whole, even though you repeat the system prompt. The same 1,050,000 tokens cost $8.40 in one request or $4.20 in 100K chunks — exactly half. A 2,000-token system prompt repeated across those chunks adds about 20,000 tokens, roughly $0.08 at promotional rates, which does not come close to cancelling the saving.
The trade-off is real, though: chunked requests cannot see across chunk boundaries. Use chunking for extraction, classification and per-section summarization; pay the long-context premium when the answer genuinely requires holding the whole document at once. 100K tokens is a convenient chunk size because it stays well under the threshold and still fits the smallest tracked window.
A window is a ceiling, not a target.
Three things get worse as you approach 100% utilization, and none of them show up on a rate card.
Cost scales linearly, value does not. Every token you send is billed, whether or not the model needed it. Filling a 1,050,000-token window on Sol costs $8.40 per request — at 1,000 requests a day that is $252,000 a month, before a single output token.
Retrieval quality decays with length. Material buried in the middle of a very long context is recalled less reliably than material at the beginning or end. Feeding 50 relevant chunks usually outperforms feeding 500 mostly-irrelevant ones, and costs a fraction as much.
Latency and failure modes grow. Time-to-first-token rises with prompt length, and a request that is close to the limit has no room for tool output, retry context or a longer-than-expected answer. Leave operational headroom — check real documents with the Context Window Checker before you batch.
Context-window questions
What is a context window in simple terms?
It is the maximum number of tokens a model can hold for one request, counted across everything you send plus everything it generates. Think of it as a single budget shared by your prompt, system instructions, retrieved documents, chat history and the reply — typically 200,000 to 1,050,000 tokens on current frontier models.
Does the context window include the response?
Yes. Input and output draw on the same budget, and most providers also set a separate, lower cap on a single response. GPT-5.6 and Claude Opus/Sonnet 5 allow up to 128,000 output tokens, Haiku 4.5 up to 64,000, and both Gemini models up to 65,536 — so a 200,000-token Haiku 4.5 window can take at most 136,000 input tokens if you want a full-length answer.
How much does it cost to fill a context window?
Filling a window entirely with input costs $0.20 on Claude Haiku 4.5, $0.3146 on Gemini 3.5 Flash-Lite, $0.42 on GPT-5.6 Luna, $0.7864 on Gemini 3.6 Flash, $2.00 on Claude Sonnet 5, $4.20 on GPT-5.6 Terra, $5.00 on Claude Opus 5 and $8.40 on GPT-5.6 Sol. The 42x spread comes mostly from window size, not just per-token rate.
What happens if I exceed the context window?
The provider rejects the request instead of answering from part of it — you get an error, not a silently degraded result. Frameworks and middleware may truncate before sending, which is worse because it looks like it worked. Count tokens before dispatch and keep operational headroom for system messages, tool schemas and retrieved context.
Is a bigger context window always better?
No. Cost scales linearly with what you put in, so a full 1,050,000-token GPT-5.6 request at the long-context rate costs $8.40 on Sol versus $0.42 on Luna for the same tokens. Latency grows with prompt length, retrieval quality tends to degrade for material in the middle of very long contexts, and OpenAI doubles input rates above 272,000 tokens. Retrieval that feeds the model only what it needs usually beats filling the window.
Where these numbers come from.
Context windows and maximum output limits are provider-published values, re-verified weekly against official pricing and model documentation. Word and page figures are planning estimates at 0.75 English words per token and 500 words per page — they are not tokenizer output, and the same text meters several percent differently across providers. Cost figures multiply token counts by published per-1M input rates; OpenAI long-context rates are applied automatically above 272,000 input tokens. Caching, batch, tool use and taxes are excluded throughout.
Official sources: OpenAI model docs, Anthropic pricing, Google Gemini API pricing. Always confirm the invoice before making purchasing decisions.