Context windows have grown from a few thousand tokens to over a million. DeepSeek V4.1 Flash and GLM 5.3 on BroadNet Inference accept 1,048,576 tokens per request, prompt and answer together. That changes how you can build, but it doesn’t make every prompt better.
What fits in a million tokens?
As a rough guide: 1,000 words of English prose came to about 1,290 tokens, averaged across five tokenizers, in our tokenizer efficiency dataset, so a million tokens is roughly 810,000 words of plain prose:
| Content | Roughly |
|---|---|
| A mid-sized codebase (tens of thousands of lines) | Fits, with room to spare |
| Several long contracts with schedules | Fits |
| A day of meeting transcripts | Fits |
| A large monorepo with dependencies | Often doesn’t; select directories |
Token counts depend on language and content. Code and Arabic text often use more tokens per character than English prose. Measure with the usage block on a real request rather than estimating, or check a sample in the context window calculator.
Which BroadNet models have a 1M-token context window?
| Model | Context window | Input | Output | Images |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | 1,048,576 | $0.175 | $0.52 | Yes |
| GLM 5.3 | 1,048,576 | $0.78 | $3.46 | No |
| Qwen3.8 27B | 262,144 | $0.175 | $2.17 | Yes |
| BroadNet Gemma 4 VL 26B | 32,768 | $0.10 | $0.22 | Yes |
Prices are per 1M tokens. GLM 5.3 adds tool calling and structured output on the same 1M window.
When is a 1M-token window worth it?
- Cross-file reasoning. “Why does this test fail?” often depends on five files nobody would think to retrieve together.
- Whole-document review. Finding conflicts between clause 4 and schedule B needs both in view.
- One-off analysis. For a question you’ll ask once, building a retrieval index costs more effort than it saves.
- Agents. A sub-agent that can read the whole module makes fewer wrong guesses. See running parallel agents.
When is retrieval still the better choice?
Retrieval wins for many small lookups over a library far bigger than any window, for content that changes every minute, and for fast interactive answers where reading a huge prompt first adds delay. Many production systems do both: retrieve the most relevant 200,000 tokens, then let a long-context model reason across all of them. The full decision, with cost per query, latency and recall side by side, is in long context vs RAG.
Does the model really use the whole window?
Not evenly. The research is consistent on this:
- The Lost in the Middle paper (Liu et al.) found that performance “is often highest when relevant information occurs at the beginning or end of the input context”.
- Chroma’s Context Rot report tested 18 models and found that “performance grows increasingly unreliable as input length grows”, even on simple tasks.
These studies did not test the models on BroadNet, so treat them as a warning, not a measurement. The practical rule: a window that accepts a million tokens does not mean every token gets equal attention. Put the most important material near the start or the end, and test recall on your own documents.
What does a million-token prompt cost?
On a pay-as-you-go key, DeepSeek V4.1 Flash bills $0.175 per 1M input tokens and $0.0055 per 1M cached input tokens. A full million-token prompt costs about $0.175 the first time, and much less on repeat calls when the prefix is cached. How to raise your cache rate.
The prompt caching calculator shows how much a cache hit saves on your own prompt sizes. A worked example: ten questions about the same 800,000-token contract set. Without caching, that is 8M tokens at $0.175 per 1M. If the documents come first and stay identical, the first call reads 800,000 new tokens and the next nine can read most of them from cache at $0.0055 per 1M. Only the question and the answer change.
On an unlimited plan the size of the prompt does not change the bill: Full Moon is $60 a month per session with no token meter. That is why long-context agents are a common reason to choose a flat plan over per-token billing.
How much room should you leave for the answer?
The window covers prompt and output, and on reasoning models the output includes thinking. If you send 1,040,000 tokens of prompt, only about 8,000 remain for reasoning and answer. Keep a margin, set max_tokens explicitly, and read why reasoning models can return an empty answer.
How to prepare a codebase for a 1M-token prompt
- Exclude what the model doesn’t need. Lock files, build output, vendored dependencies, images and generated code take tokens and add noise.
- Count before you send. Send a small sample, read
prompt_tokensfrom usage, and scale up. - Order by stability. Rarely changed files first, the files under edit last, the question at the very end. That keeps the prefix cacheable.
- Label every file. Wrap files in clear markers such as
=== path/to/file.py ===so the model can cite them. - Ask for citations. “Quote the line you rely on” reduces confident mistakes over huge inputs.
- Measure. Log prompt tokens, cached tokens and time to first token per request. Streaming shows the first token as soon as it arrives.
Where BroadNet Inference fits
Two of its chat models take 1,048,576 tokens per request, and BroadNet does not trim your context. On a per-token key, caching keeps repeated long prompts affordable; on a plan session, prompt size does not affect the bill. If your corpus is far larger than a million tokens or changes every minute, you still need retrieval, and no window size changes that.

