Loading prices...
All news
Glossy violet and silver window frame glowing from inside, a long white paper scroll flowing through it with only the part inside the frame lit, black background with violet glow

Context window in AI: GPT-6.2, Claude and Gemini all hold about a million tokens

17:00 · 20.11.2026
Source: Anthropic
3

A context window is how much text a model can hold at once: your request, the conversation so far, any files you attach and the answer it writes. The three largest labs have now converged on roughly the same number for their top models. Our AI glossary explains the other terms around them.

What the models hold

  • OpenAI: GPT-6.2 Astra and GPT-6.2 Sol take 1.05 million tokens of context and write up to 128,000 tokens in one answer
  • Anthropic: Claude Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 take 1 million tokens and write up to 128,000
  • Google: Gemini 4 Carbon, Argon and Hydrogen take 1 million tokens

The figures come from each company's developer documentation on 18 November. A window is measured in tokens, the chunks of text a model reads. When we counted tokens in our own news with OpenAI's tokenizer, English came to about 1.4 tokens per word and Russian to about 1.8. By that measure a million tokens holds roughly 700,000 English words or 540,000 Russian ones, several long novels at once. Other vendors' tokenizers will give somewhat different numbers.

Where long context falls short

A model that accepts a million tokens does not use all of them equally well. Researchers at Stanford and other universities tested this in 2023 by moving the one useful document around inside a long prompt.

“Performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts.”

— Nelson F. Liu and co-authors, Lost in the Middle: How Language Models Use Long Contexts

Quote source: arXiv, 2023. The study tested models of that year, and current models may do better.

Long prompts also cost time and money. You pay for input tokens on each request, though many providers bill repeated text at a lower cached rate, and a full window makes the model slower to start answering. Anthropic said in November that its new message threads cut the median wait for the first token by about 35% on requests of 500,000 to 1 million tokens, as we reported.

Our take

Maks, the Intokened editor, says LLM was the hardest AI term for him to grasp. The context window is the part of an LLM that ordinary users run into first: the moment a long chat starts forgetting what you said at the beginning, you have filled it.

Put the instruction and the key document at the start or the end of your prompt, not in the middle. Start a fresh chat when a long one drifts. For a large archive, RAG that pulls only the relevant pages into the window usually beats pasting everything in.

Context sizes come from the developer documentation of OpenAI, Anthropic and Google on 18 November 2026. Word estimates use our own count with the o200k_base tokenizer and are approximate.

Published: 17:00 · 20.11.2026
Maks

Author

Maks

Trading man

I've been interested in the cryptocurrency market for a long time, am a trader, and write articles and news about my experience and crypto in simple terms.

Comments (0)

No comments yet — be the first!

Join our community

No spam, unsubscribe in one click