Biggest context windows, and the price of using them
A model's context window caps how much you can feed it in one call — the whole point of long-context work like RAG, big-codebase reasoning, or long documents. Across 208 models here, the median window is 256k tokens and the largest reaches 1.31M. The window is not fixed by the weights alone: each host decides how much of it to serve, and 49 of the 112 models sold by several hosts here (44%) come with different windows depending on where you buy them. Every figure below is the largest window a host actually serves, priced at the cheapest host that serves it.
Cheapest long context: qwen3.7-flash gives you a 1M-token window at just $0.16/1M blended (Alibaba).
Long windows aren't free to use: cost scales with tokens sent, and the output premium still applies — see output vs input and prompt caching for how to keep a big-context bill down.
How many models reach each size
Share of models whose context window is at least this large.
Does a bigger window cost more?
Every model plotted by its largest served window against the cheapest blended price at that window (both log scales). Less correlated than you'd expect: million-token windows exist at budget prices, and some small-window flagships are the dearest dots here. Click a brand to isolate it.
Largest context windows
Ranked by the largest window any host actually serves, priced at the cheapest host that serves that window — so the size and the price on a line describe the same endpoint. The window is a deployment choice, not a property of the weights: where hosts of one model disagree, the line is marked and the smaller figure is named.
| Model | Context | Max output | Cheapest host at that window | Blended /1M |
|---|---|---|---|---|
| llama-4-scout | 1.31M ✳ | 8,192 | $0.95 | |
| mimo-v2.6-flash | 1M | 131,072 | Relace | $0.28 |
| mimo-v2.5 | 1M | 131,072 | DeepInfra | $0.399 |
| gpt-6-luna | 1M | 128,000 | OpenAI | $0.6 |
| mimo-v2.5-pro | 1M ✳ | — | GMICloud | $0.913 |
| mimo-v2.6-pro | 1M | — | DeepInfra | $1.17 |
| gpt-5.6-luna | 1M | 128,000 | OpenAI | $1.4 |
| gpt-6-sol | 1M | 128,000 | OpenAI | $12 |
| gpt-6.1-sol | 1M | 128,000 | OpenAI | $12 |
| gpt-5.6-terra | 1M | 128,000 | OpenAI | $14 |
| gpt-5.6-sol | 1M | 128,000 | OpenAI | $24 |
| glm-5.3-flash | 1M ✳ | 128,000 | StreamLake | $0.189 |
| deepseek-v4-flash | 1M ✳ | — | Relace | $0.21 |
| deepseek-v4-flash-0423 | 1M ✳ | 384,000 | StreamLake | $0.21 |
| deepseek-v4.1-flash-1017 | 1M | 131,072 | DeepInfra | $0.392 |
| deepseek-v4.1-flash | 1M | 131,072 | Io Net | $0.468 |
| minimax-m3.1-flash | 1M | 524,288 | InferenceNet | $0.6 |
| nemotron-3.5-super-vl | 1M ✳ | — | CoreWeave | $0.65 |
| deepseek-v4-flash-vision-exp | 1M | 262,144 | DeepInfra | $0.862 |
| llama-4-maverick | 1M ✳ | — | DeepInfra | $1 |
✳ not every host of that model serves that window — the model page lists which ones do. Context in tokens; prices per 1M tokens (USD). Verified 2026-10-23.
FAQ
What is the largest LLM context window available?
The largest window in this index reaches 1.31M tokens, against a median of 256k across 208 models. Window size caps how much you can send in a single call.
What's the cheapest way to get a long context window?
qwen3.7-flash offers a 1M-token window at just $0.16 per 1M tokens blended via Alibaba — the cheapest long-context option here.
Does the provider change a model's context window?
Yes, often. The weights allow a maximum window, but each host decides how much of it to serve. Of the 112 models here sold by more than one host, 49 (44%) are served with different windows depending on the host — qwen3.8-27b runs from 66k to 1M, a 15.3× gap on identical weights. So routing to the cheapest host can silently cost you context length: check the window on the endpoint you actually call, not on the model.