Context windows explained - 128K vs 200K vs 1M tokens

The "context window" is one of the most quoted specs in AI, yet one of the least understood. A model advertised with a 1M-token window sounds eight times more capable than a 128K one. In practice the difference is more nuanced. Here is what the numbers actually mean and when a bigger window earns its keep.

What a context window actually is

The context window is the maximum number of tokens a model can consider at once. Crucially, it is shared: it covers your system prompt, the full conversation history, any files or retrieved documents you paste in, and the answer the model generates. If the input fills the window, there is no room left to reply.

A token is roughly 3-4 characters of English, so about 0.75 words. Quick conversions:

These are ceilings, not free space. Reserve a few thousand tokens for the output, and remember that long histories silently eat the budget every turn.

128K vs 200K vs 1M: when each makes sense

128K is the modern baseline and handles the vast majority of work: chat, drafting, summarizing a report, refactoring a file, answering questions over a handful of documents. If your task fits in a few dozen pages, 128K is plenty and usually cheaper and faster.

200K buys breathing room for longer sessions — extended agentic coding runs, large diffs, or analyzing a sizable contract without aggressive trimming. The jump from 128K to 200K is incremental, not transformational.

1M is a different category. It enables dropping an entire repository, a full deposition, or hundreds of pages of research into a single prompt and reasoning across all of it. The trade-offs are real: latency rises, cost scales with tokens, and most models show "lost in the middle" behavior, where facts buried in the center of a huge prompt get less attention than those at the start or end.

Bigger isn't automatically better

A large window is not a substitute for good retrieval. Stuffing 900K tokens of documents into context to answer one question is slow and expensive, and accuracy can drop. Often a focused 128K prompt with the right 10 pages beats a 1M prompt with everything. Use the big window when relationships across the whole corpus matter — cross-referencing, global refactors, full-document consistency checks — not as a dumping ground.

A practical rule: pick the smallest window that comfortably holds your real task plus a margin for the reply.

Switch windows without switching stacks

Different models ship different limits, and you rarely know upfront which fits a job. On AnyModel every model lives behind one OpenAI-compatible endpoint, so testing a 128K model against a 1M one means changing a single string — the model id — not your SDK, keys, or base URL.

curl https://anymodel.org/v1/chat/completions \
  -H "Authorization: Bearer $ANYMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gemini-2.5-pro","messages":[{"role":"user","content":"Summarize this repo."}]}'

Want a smaller, cheaper window for the same call? Swap gemini-2.5-pro for a Claude, GPT, DeepSeek, or Qwen id and rerun. One key reaches GPT, Claude, Gemini, DeepSeek, GLM, Kimi, Qwen, and Grok — see the live model comparison to match window size against price.

Using a CLI? One line wires it up:

bash <(curl -fsSL "https://anymodel.org/i?tool=codex") <YOUR_API_KEY>

(Use tool=claude for Claude Code. For Cursor, Zed, Cline, Aider, or Continue, point the OpenAI base URL at https://anymodel.org/v1 with your key.)

Privacy on long prompts

Large context often means pasting sensitive material. With Ghost Mode you can opt into zero-retention keys — we don't store prompts or responses, only a token counter. Note the honest caveat: the model provider still receives the prompt, so this isn't absolute privacy, but nothing lingers on our side.

Try it

Start free with 1,000,000 tokens, or 6,000,000 once you link Telegram — no credit card, no subscription. Test 128K, 200K, and 1M models side by side and let your task pick the window. Create an account and run your first prompt in a minute. More guides on the blog.