You have a 600 page contract, a year of support tickets, or an entire codebase, and you want an AI model to read all of it before answering. Two years ago that meant chopping the material into pieces and hoping nothing important fell between them. Today, several models accept a million tokens or more in a single prompt. The catch is that the headline number rarely tells you what you will actually get. This article sorts out the best long context options, the real limits behind the big numbers, what you can use for free, and what a full window costs when the invoice arrives.
💡 Quick answer: For most people, a 1M token window from Gemini, Claude or GPT is plenty. Llama 4 Scout advertises 10M, but accuracy past a few hundred thousand tokens depends far more on the task than on the number printed in the spec sheet.
What a Context Window Really Is
The context window is the working memory of a language model. Everything it can "see" while writing an answer has to fit inside it: your instructions, the documents you paste in, the earlier turns of the chat, and the reply it is generating. Once the budget is gone, older material gets dropped or the request fails.
Tokens, Not Words
Models count tokens, which are chunks of text roughly three quarters of an English word long. A few reference points make the numbers easier to picture:
| Window size | Approximate words | What fits |
|---|
| 8K tokens | 6,000 | A long article |
| 128K tokens | 96,000 | A full novel |
| 200K tokens | 150,000 | A thick technical manual |
| 1M tokens | 750,000 | Several novels or a mid sized codebase |
| 10M tokens | 7,500,000 | A small library (advertised only) |
Code, tables and non English text usually burn more tokens per word, so a 1M window holds less than the table suggests when your material is dense.

Input and Output Share the Budget
A common surprise: the window is shared. If a model has a 1M window and you send 990K tokens, only about 10K are left for the answer. Many models also set a separate maximum output (often 64K to 128K tokens), so a huge window does not mean a huge reply. When you plan a job such as "rewrite this whole manuscript," check the output cap first, not only the input size.
Best Long Context Options Right Now
The market has moved quickly. A mid 2026 comparison counted well over a dozen models with a window of 1M tokens or more, and the club keeps growing. Here is how the options available on PicassoIA stack up by role.

The One Million Token Group
These are the models most people should look at first when a document is too big for a normal chat:
- Gemini 3.1 Pro: Google has pushed long context harder than anyone, with windows of 1M tokens and, in some configurations, 2M. A strong pick for video transcripts, big PDFs and mixed text and image inputs.
- Gemini 3.5 Flash: the faster, cheaper sibling. Good when you need to scan a lot of text quickly and can accept slightly less depth.
- Claude Opus 4.7: reported to offer a 1M window at its standard per token price, with no extra long prompt surcharge. That matters for budgets.
- Claude Sonnet 5 and Claude Fable 5: strong at following long, detailed instructions across a large body of material, which is where many models slip.
- GPT 5.6 Luna, GPT 5.6 Terra and GPT 5.6 Sol: OpenAI's current line. One listing describes Luna with a context window of roughly 1.05M tokens and up to 128K tokens of output.
💡 Window sizes change with every release and sometimes vary by access tier. Treat the numbers here as a snapshot and check the model page before you plan a big job.
Llama 4 Scout and the 10M Claim
Llama 4 Scout Instruct holds the record for the biggest advertised window: 10 million tokens. That is an impressive engineering number, but hosted versions often cap the usable length well below it. Think of 10M as proof that the architecture can stretch, not as a promise that you can drop ten novels in and get perfect recall. Its sibling Llama 4 Maverick Instruct is worth testing when you want an open family model for general chat.

Budget and Open Picks
Not every job needs a frontier model. Cheaper options keep showing up in long context rankings:
- Qwen3.7 Plus: a capable generalist that also reads images.
- Kimi K2.6: popular for agent style work and code across many files.
- Granite 4.0 H Small: listed as a free long context model, built on a hybrid design that keeps memory use low.
- Granite 8B Code Instruct 128K: a 128K window tuned for code. Small, quick and plenty for one repository.
- DeepSeek V3: a low cost workhorse for long text jobs where top tier reasoning is not required.
DeepSeek V4 Flash has been cited as one of the cheapest ways to run a full 1M token request, at around fourteen cents. It is not on PicassoIA yet, so for a similar low cost approach today, try DeepSeek V3 or Qwen3.7 Plus above.
Real Limits Behind the Big Numbers
A model that accepts a million tokens is not the same as a model that uses a million tokens well. This is the part spec sheets skip.
Advertised vs Effective Length
Independent long context benchmarks keep finding the same pattern: retrieval of a single obvious fact stays strong across the window, but tasks that need several facts spread through the material degrade much sooner. One roundup put the gap between advertised and effective length at 30 to 60 points for multi fact retrieval past 200K tokens. In plain terms, you can paste in the whole archive, but the answer to a tricky cross reference question may only be reliable for the first few hundred thousand tokens.

Lost in the Middle and Other Traps
Four limits show up again and again:
- Middle blindness. Models pay the most attention to the start and the end of the prompt. Facts buried in the middle get missed more often.
- Latency. A full 1M token prompt can take well over a minute to process before the first word appears.
- Output caps. As noted above, a giant input does not buy you a giant reply.
- Distraction. Irrelevant text is not harmless. Adding unrelated documents can lower accuracy on the part you care about.
💡 Rule of thumb: If the answer lives in one or two places, a smaller window plus good retrieval beats a giant paste. If the answer depends on the whole document at once, such as tracking a contradiction across a contract, go big.
Free Use: What You Actually Get
"Free" and "long context" rarely appear in the same sentence, because processing a million tokens costs real compute. There are still sensible ways to experiment without paying much.

Free Tiers and Their Catches
Most providers run limited free access to their chat apps or API, but the long window is usually the first thing they restrict. Typical patterns:
- A free chat tier with a smaller window than the paid one.
- A free API tier with tight rate limits that a single big prompt can burn through.
- Trial credits that disappear fast when each request is hundreds of thousands of tokens.
If you test a free tier, start with a 50K token sample of your material. You will learn whether the model handles your content type before you spend anything on a full run.
Free and Open Models
Open weight models are the other route. Several on PicassoIA are labeled as free to try, including Granite 4.0 H Small, Granite 4.1 8B and Llama 4 Maverick Instruct. Their windows are smaller than the frontier giants, usually in the 128K range, but that already fits a full book, a long report or a medium sized repository. For many everyday jobs, that is the sweet spot between cost and capability.
| Need | Sensible free route |
|---|
| Summarize one long report | A 128K open model such as Granite 4.0 H Small |
| Review a single repository | Granite 8B Code Instruct 128K |
| Casual long chat | Llama 4 Maverick Instruct |
| Mixed text and image documents | Qwen3.7 Plus or a Gemini Flash model |
Pricing Without Surprises
Context pricing is simple in principle: you pay per million input tokens and per million output tokens. The trouble is that long prompts multiply the input bill quickly, especially in chat apps where the whole history is resent on every turn.

The Cost of One Full Window
A quick calculation shows why this matters. At an input price of $5 per million tokens, one full 1M token prompt costs about $5 before the model writes a single word. At $1.25 per million it is about $1.25. At fourteen cents, it is fourteen cents. Ask ten follow up questions on the same full prompt without caching and you pay for that window ten times.
| Scenario | Input price per 1M tokens | One full 1M prompt | Ten follow ups, no caching |
|---|
| Premium frontier model | about $5 | about $5 | about $50 |
| Mid priced model | about $1.25 | about $1.25 | about $12.50 |
| Budget model | about $0.14 | about $0.14 | about $1.40 |
These figures are examples drawn from recent public price listings. Prices move often, so check the current rate before you commit to a large batch.
Caching and Surcharges
Two details change the math more than the headline rate:
- Prompt caching. Many providers charge a fraction of the normal rate for input you reuse. If you ask many questions about the same document, caching can cut the bill dramatically.
- Long prompt surcharges. Some vendors, including Google, charge a higher rate once a prompt passes a threshold. Anthropic has been reported to have dropped its long context surcharge on recent models, which makes cost easier to predict.
💡 Budget habit: Estimate tokens first (words times 1.3), multiply by the model's input price, then multiply by the number of questions you plan to ask. If the result makes you wince, trim the material or switch to a cheaper model.
Long Context or RAG?
Retrieval augmented generation, or RAG, means searching your documents first and sending only the relevant passages to the model. It was the standard workaround when windows were small, and it is still a serious competitor.

| Factor | Long context | RAG |
|---|
| Setup effort | Almost none, just paste | Needs indexing and a search layer |
| Cost per question | High on big inputs | Low, small prompts |
| Cross document reasoning | Strong when it fits | Weaker, pieces can be missed |
| Freshness | Re-paste each time | Index updates automatically |
| Best for | One off deep reads | Repeated questions on a huge, changing library |
Use long context when you need one deep read of a bounded set of files, such as a contract bundle, a research paper set or a single repository. Use RAG when the library is enormous, changes daily, or you will ask thousands of small questions. Many strong setups combine both: retrieve a few hundred thousand tokens of the best candidates, then let a long window model reason across them.
A big window rewards a little structure. These habits improve results on almost every model.

Structure the Prompt
- Put the question after the documents. Place your instruction at the very end, and repeat the key request in one line if the material is long.
- Label every document. Use simple tags such as
<doc id="3" title="Lease 2024"> so the model can cite where a fact came from.
- Order by importance. Since the start and end get the most attention, put the most critical material there.
- Remove noise. Strip headers, footers, duplicate pages and boilerplate before pasting.
Verify the Answer
- Ask for quotes. Request the exact passage that supports each claim, then spot check three or four of them.
- Run a needle test. Hide a made up fact in your material and ask about it. If the model misses it, lower your trust for that length.
- Split when in doubt. Two prompts of 300K tokens often beat one of 600K for accuracy.
Put a Big Window to Work
The best way to find out where a model's memory starts to fray is to test it on your own material. Pick a document you know well, paste it into a model such as Gemini 3.1 Pro, Claude Opus 4.7 or Llama 4 Scout Instruct, and ask three questions: one easy, one that needs two distant facts, and one that checks a detail from the middle. Then repeat with a cheaper option and compare.

Here is a simple routine that works on PicassoIA:
- Open the Large Language Models collection and choose a model from the lists above.
- Paste a 50K token sample first to check quality before a full run.
- Put your question last, label each document, and ask for supporting quotes.
- Compare two models on the same prompt and keep the cheaper one if the answers match.
Once your long documents are summarized, PicassoIA can also turn the results into visuals, voiceovers and video clips. Try it on your own text today, and see how far a bigger window can take your next project.