The largest context window AI models — November 2026, ranked by window size.
The largest context window AI models are no longer rare: 38 of the 59 models in our database accept at least a million tokens in one request. This post ranks the biggest windows, names the single 10-million-token outlier, prices what it costs to actually fill a million-token window at each model's own input rate, and explains why window size is not the same thing as output size.
By AI Pro League · Published
Largest context window AI models, from our AI model rankings database: the outright leader is Llama 4 Scout at 10,000,000 tokens — roughly ten times the next tier — but it carries no League Score and no API price in our records, so it tops the size table and nothing else. The highest-scored million-token model is GPT-6 Astra at 99.8. The full long-form shortlist lives in our best AI models for long documents picks, and every window below can be checked on the model’s own page in the model database.
The largest context window AI models, ranked
| # | Model | Maker | Context (tokens) | League Score | $/1M in |
|---|---|---|---|---|---|
| 1 | Llama 4 Scout | Meta | 10,000,000 | — | — |
| 2 | GPT-6 Astra | OpenAI | 1,050,000 | 99.8 | $10.00 |
| 3 | GPT-6.1 Sol | OpenAI | 1,050,000 | 96.6 | $2.00 |
| 4 | GPT-5.6 Sol | OpenAI | 1,050,000 | 86.7 | $4.00 |
| 5 | GPT-6.1 Terra | OpenAI | 1,050,000 | 83.2 | $1.00 |
| 6 | GPT-5.6 Terra | OpenAI | 1,050,000 | 68.6 | $2.00 |
| 7 | GPT-5.6 Luna | OpenAI | 1,050,000 | 66.7 | $0.20 |
| 8 | GPT-6 Luna | OpenAI | 1,050,000 | 55.1 | $0.10 |
| 9 | Gemini 4 Argon | Google DeepMind | 1,048,576 | 86.4 | $2.00 |
| 10 | Muse Wonder | Meta | 1,048,576 | 75.5 | $4.00 |
Read the table’s shape, not just its order: one outlier at 10,000,000 tokens, then a dense plateau. 38 models sit at or above 1,000,000 tokens — 64% of the whole database — and the median window across all 59 models is 1,000,000 tokens. A million tokens has become the standard flagship window, not the exception.
How many AI models have a 1 million token context window?
38 of the 59 models we track — and they are not all flagships. The million-token club includes the overall leader, GPT-6 Astra (1,050,000 tokens, League Score 99.8), and also the cheapest input rate in the club: GPT-6 Luna at $0.10 per million input tokens. Window size no longer predicts price or rank — which is exactly why the worked example below matters more than the raw token count.
What filling a 1M window actually costs — the worked example
A context window is only useful if you can afford to fill it. Price one full window at each model’s own input rate from our AI model pricing index (GPT-6 Astra: $10.00 per 1M over a 1,050,000-token window; GPT-6 Luna: $0.10 per 1M over a 1,050,000-token window):
| Model | Window | $/1M in | Fill cost |
|---|---|---|---|
| GPT-6 Astra (overall #1) | 1,050,000 | $10.00 | $10.50 |
| GPT-6 Luna (cheapest 1M input) | 1,050,000 | $0.10 | $0.10 |
Filling the leader’s window costs about 100× filling the cheapest million-token window — before output is billed at all. Repeated-context work narrows that gap: both models publish cached-input rates, and our prompt caching pricing explainer prices the discount across the field.
Context window is not output length
The window counts everything in the request — your prompt plus the reply share one budget on most APIs, and the reply has its own, much smaller cap. Most million-token models in our records cap output at 128,000 tokens or below; the widest recorded output cap in the club belongs to Gemini 4 Argon, whose maximum output equals its 1,048,576-token window. A model that can read a million tokens cannot necessarily write a long document in one reply — check both numbers on its model page before planning around either.
Does a bigger context window mean a better model?
Not in our data. The 10-million-token leader is unscored, several million-token models sit in the bottom half of the League table, and the median window (1,000,000 tokens) already holds more text than most real documents. Window size answers "will it fit?" — the benchmark components behind the League Score answer "how well will it do once it fits?", and our explainers for LiveBench and SWE-bench Verified set out what those scores do and do not prove.