How Context Windows Power Smarter AI Answers in Your Browser
The context window is the most important number in AI assistants — it determines how much of a page your browser extension can actually analyze, and why a bigger window means fewer frustrating cutoffs.
Key takeaways
- A context window is the maximum amount of text an AI can read and respond to in one request — measured in tokens, roughly 0.75 words each.
- Gemini 1.5 Flash supports up to 1 million tokens, making it practical to analyze entire research papers or long web pages in a single pass.
- Browser extensions that only send a page summary to the AI can miss crucial details in the sections they cut; sending the full text produces better answers.
- Token cost scales with context size, so BYOK tools that use Gemini's free tier let you use large contexts without a per-request charge.
On this page
If you have ever asked an AI browser assistant a question about a long article and received an answer that completely ignored something written on the page, the most likely culprit is the context window — specifically, the fact that whoever built the extension decided to send only a portion of the page text to the model.
Understanding what context windows are, how they are measured, and how they differ between models explains a lot about why two AI tools can behave so differently on the same piece of content.
What a context window actually is
A context window is the total amount of text — your question, the page content, any instructions the tool bakes in, and the model’s answer — that a language model can hold in “working memory” during a single request. Everything inside the window is visible to the model simultaneously. Everything outside it is invisible.
The unit of measurement is tokens, not words or characters. A token is roughly four characters of English text, so one word is approximately 0.75 tokens. “Context window” is often stated in tokens, as in “a 128k context” — meaning roughly 128,000 tokens, or about 96,000 English words.
This matters for browser extensions because a web page can be long. A detailed Wikipedia article runs 5,000–15,000 words. A research paper might be 10,000 words. An in-depth news analysis or technical documentation page can exceed that. If your extension caps out at a small context window and the page is longer than that cap, it has to make a decision: truncate the page, summarize it, or ask the model to answer without reading all of it.
How different models compare
Models differ substantially in context size. Older or smaller models have windows in the tens of thousands of tokens. Modern frontier models have pushed much higher:
- Gemini 1.5 Flash supports up to 1 million tokens
- Gemini 1.5 Pro supports up to 2 million tokens
- Gemini 2.0 Flash supports up to 1 million tokens
- GPT-4o and Claude 3.5 Sonnet each support 128k tokens (roughly 96,000 words)
A million-token context window means the entire text of a 700-page novel, or a full semester’s worth of lecture transcripts, fits in a single request. For typical web pages, it effectively removes the limit — a 1M token context window will never truncate a web article.
Why this matters for a browser extension specifically
When you open a page and ask your AI assistant about it, the extension has to decide what to include in the API request. The straightforward approach is: send the whole page. The cheaper, faster approach is: extract a summary or send only a portion.
The problem with the second approach is precision. AI models answer questions based on what is in their context. If the passage that answers your question was cut to save tokens or reduce API cost, the model will either guess, refuse, or hallucinate an answer — none of which are useful. The frustrating “I don’t see that information on this page” response from an assistant is very often the result of that information being in the 40% of the page that was left out of the request.
A 1M-token context window makes the cautious approach unnecessary. There is no reason to truncate the average web page when the limit is so far above what a single page contains.
Why BYOK makes large contexts practical
API costs scale with the number of input tokens in a request. A 10,000-word page sent to a model with standard pricing adds up, especially across dozens of queries per day.
This is one of the reasons why bring-your-own-key tools — where you supply your own API key and are billed directly by the provider — have a structural advantage for heavy users. Gemini offers a free tier at limited request volumes, and its Flash models are among the least expensive per token even at paid rates. When you are not paying per query through an intermediary, using a large context becomes the default rather than a premium feature.
What happens at the token limit
Even with a large window, hitting the limit has well-defined consequences. If the combined length of the system prompt, page content, conversation history, and your question exceeds the model’s context size, the API returns an error — or the extension silently truncates input to fit.
Most well-designed extensions handle this gracefully: if the full page text plus conversation history would exceed the window, they trim the oldest conversation turns first, then fall back to partial page content, and surface a warning that the page is too long to analyze completely. This is more honest than silently cutting the page and answering as if the full text were available.
The practical summary
Context window size is not a marketing number. It directly determines whether your browser AI can answer questions about the full text of a page or only about an arbitrarily selected portion of it. For everyday web reading and research, a model with a million-token context window paired with a tool that actually sends the full page text to it is meaningfully more useful than the same question sent to a model that truncates at 10,000 tokens — regardless of which underlying model is considered more capable in benchmarks.