The typical new AI model now reads a million tokens
In 2024 the middle-of-the-pack new model could take in about 131,000 tokens at once. In 2026 it takes a million. Most of that jump happened in the last six months.
Claude 3 Opus arrived in March 2024 with a 200,000-token context window, and that was large for the time. Of the 42 language models in our catalogue from 2024, 31 had a smaller one.
We have added 111 language models so far in 2026 and 69 of them accept a million tokens or more. The context window is the amount of text a model can hold in view at once: your question, any documents you attach, the conversation so far and its own answer.
The numbers, year by year
We took every language model in the catalogue that has a release date and a context window, removed entries whose names mark them as dated snapshots or fast tiers, and found the median for each year. The median is the middle value: half the models are above it and half below. It suits this job better than an average, because one model with a four-million-token window would drag an average upwards on its own.
Median context window of new language models, by year
From 2024 to 2025 the median went from 131,072 to 200,000, a rise of about half. From 2025 to 2026 it went up five times over.
Slow, then sudden
Split the same models into half-years and you can see when it happened. For two years the median rose steadily. Then, between the first and second halves of 2026, it tripled.
Median context window of new language models, by half-year
| Period | New models | Median window | Under 200,000 | A million or more |
|---|---|---|---|---|
| 2024 H1 | 11 | 128,000 | 8 (73%) | 2 (18%) |
| 2024 H2 | 31 | 131,072 | 23 (74%) | 1 (3%) |
| 2025 H1 | 48 | 200,000 | 23 (48%) | 11 (23%) |
| 2025 H2 | 49 | 262,144 | 14 (29%) | 5 (10%) |
| 2026 H1 | 60 | 336,000 | 9 (15%) | 25 (42%) |
| 2026 H2 | 51 | 1,048,576 | 3 (6%) | 44 (86%) |
The last two columns show the change better than the median does. In the second half of 2024, 23 of 31 new models had a window under 200,000 tokens and one reached a million. Since July this year it is the other way round: 3 of 51 are under 200,000 and 44 are at a million or more.
Who moved
Nearly everyone. Of the 51 models listed since July, 44 accept a million tokens or more. Here they are by lab.
| Lab | New models | A million tokens or more |
|---|---|---|
| OpenAI | 16 | 16 |
| 8 | 6 | |
| Meta | 7 | 6 |
| Anthropic | 6 | 6 |
| xAI | 3 | 0 |
| Zhipu AI | 3 | 3 |
| Alibaba | 2 | 2 |
| Moonshot AI | 2 | 2 |
| NVIDIA | 2 | 1 |
| DeepSeek | 1 | 1 |
| MiniMax | 1 | 1 |
7 models fall short of a million: Grok 4.5 (500,000), Grok 4.6 (500,000), Grok 4.7 (500,000), Nemotron 3.5 Lightning (262,144), Gemini Omni 1.1 Flash (131,072), Gemini Omni 1.2 Flash (131,072) and Muse Glimmer 30B (131,072). Going by their names, the four with the shortest windows are small or fast models, which is where you would expect one.
One caution about that 44. OpenAI accounts for 16 of the 51 entries, because each of its models also has a Pro tier that we list separately. Leave OpenAI out altogether and the count is 28 of 35, which is still 80%.
What a million tokens holds
A token is a chunk of text, usually part of a word. OpenAI's guide to tokens (read on 14 November 2026) gives the rule of thumb that 100 tokens is about 75 words of English. On that rule the 2024 median of 131,072 tokens is roughly 98,000 words, and a million tokens is roughly 750,000.
That is the difference between attaching one long report and attaching a shelf of them. It is a rule of thumb for English prose, though. Code, tables and other languages split into tokens differently, so your own files may fit more or less than the rule suggests. Our token counter will give you a real figure for a piece of text.
What the number does not tell you
The context window is a limit on how much a model will accept. It says nothing about how well the model uses what you give it.
The public test we know of is a July 2025 report from Chroma, which tested 18 models and found that "model performance degrades as input length increases", even on simple tasks. That report is older than all of this year's models, so it cannot tell us whether they have fixed the problem. We have not tested it ourselves and we do not have an independent source that has. Until one exists, treat a million-token window as room you can use, not a promise about quality when you fill it.
Filling it also costs money, because you pay for every token you send, every time you send it. The median input price among the active models we track is $1.25 per million tokens, so one full million-token request costs about that much at the median before the model has written a word. At the top of the range, GPT-5.5 Pro lists at $30 per million input tokens. A long conversation resends its history on each turn, so those costs repeat.
If you are choosing a model
If you are choosing between the large models released this autumn, the context window will not separate them. Forty-four of the last 51 share almost the same figure. Price, speed and how well the model does your actual work are better things to compare, and our guide to choosing a model walks through them.
The window still matters when you look at smaller, cheaper or older models, where 131,072 and 262,144 are common. Check the figure on the context window page before you plan a job around long documents. For how context windows work underneath, see our explainer.