An AI's context window is the text it can read and keep in view at once, counted in small chunks of words called tokens. Anthropic says 200,000 tokens comes to about 500 pages of text7.
Ai Lately checked the 20 biggest windows on Nov. 12, using each maker's own pages. Every one of them now holds 1 million tokens or more. OpenAI's GPT-6 Astra and GPT-6.1 Sol lead at 1,050,00012. The other 18 each list 1 million.
The size race has ended in a tie. What matters now is recall, the price of a long prompt, and how much of the window your chat plan hands you.
Today's comic has our robot picture a life apart from the boy. It waits at the end of this piece.
The 20 Biggest Context Windows
We ranked the 20 by window size first. Among equal windows, a model with a published long-recall score goes first, highest score on top. Next comes max output, the longest reply a model can write in one turn. A cheaper input price breaks any tie that remains.
| Rank | Model | Maker | Window | Max output |
|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 1.05 million | 128,000 |
| 2 | GPT-6.1 Sol | OpenAI | 1.05 million | 128,000 |
| 3 | Gemini 4 Carbon | 1 million | Shares the window | |
| 4 | Gemini 4 Argon | 1 million | Shares the window | |
| 5 | Claude Opus 5.5 | Anthropic | 1 million | 128K |
| 6 | DeepSeek V4.1 Flash | DeepSeek | 1 million | 384K |
| 7 | DeepSeek V4 Pro | DeepSeek | 1 million | 384K |
| 8 | LongCat-2.5 | Meituan | 1 million | 256K |
| 9 | Mercury 3 | Inception | 1 million | 131,072 |
| 10 | Qwen3.8-Max | Alibaba | 1 million | 131,072 |
| 11 | GLM-5.3 | Z.ai | 1 million | 128K |
| 12 | Claude Sonnet 5.5 | Anthropic | 1 million | 128K |
| 13 | Claude Fable 5.2 | Anthropic | 1 million | 128K |
| 14 | Nova 2 Lite | Amazon | 1 million | 65,536 |
| 15 | Muse Ember | Meta | 1 million | Unlisted |
| 16 | Mistral Large 4 | Mistral AI | 1 million | Unlisted |
| 17 | Grok 4.3 | xAI | 1 million | Unlisted |
| 18 | Muse Wonder | Meta | 1 million | Unlisted |
| 19 | Kimi K3.1 Pro | Moonshot AI | 1 million | Unlisted |
| 20 | MiniMax-M3 | MiniMax | 1 million | Unlisted |
OpenAI, Google and Anthropic list their figures on their own pages12345. The other 11 makers do the same1011121314151617181920. Google counts input and output inside one shared window. "Unlisted" means the maker's page gives the window and leaves the output cap out.
All 20 of the biggest AI context windows hold between 1 million and 1.05 million tokens
Context window in millions of tokens, as each maker's own page lists it, in Ai Lately's rank order, Nov. 12, 2026
Each bar is one model's context window; the two magenta bars are OpenAI's models at 1.05 million tokens, and the blue bars are the 18 models at 1 million.
Google counts input and output inside one shared 1 million token window; the other makers list the window for input.
Sources: OpenAI API documentation [1]; OpenAI API documentation [2]; Google DeepMind [3]; Google DeepMind [4]; Claude Platform documentation [5]; Kimi API Platform [10]; DeepSeek API Docs [11]; LongCat [12]; Alibaba Cloud Model Studio [13]; Inception documentation [14]; Z.ai documentation [15]; Amazon Nova User Guide [16]; xAI documentation [17]; Meta AI for Developers [18]; MiniMax API documentation [19]; Mistral AI documentation [20]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| GPT-6 Astra | 1.05 million |
| GPT-6.1 Sol | 1.05 million |
| Gemini 4 Carbon | 1.00 million |
| Gemini 4 Argon | 1.00 million |
| Claude Opus 5.5 | 1.00 million |
| DeepSeek V4.1 Flash | 1.00 million |
| DeepSeek V4 Pro | 1.00 million |
| LongCat-2.5 | 1.00 million |
| Mercury 3 | 1.00 million |
| Qwen3.8-Max | 1.00 million |
| GLM-5.3 | 1.00 million |
| Claude Sonnet 5.5 | 1.00 million |
| Claude Fable 5.2 | 1.00 million |
| Nova 2 Lite | 1.00 million |
| Muse Ember | 1.00 million |
| Mistral Large 4 | 1.00 million |
| Grok 4.3 | 1.00 million |
| Muse Wonder | 1.00 million |
| Kimi K3.1 Pro | 1.00 million |
| MiniMax-M3 | 1.00 million |
Bigger Windows, Weaker Recall
A window tells you what fits, while recall tells you what the model can still find and use. Those two numbers drift apart as a prompt grows longer.
Google DeepMind measured this on its Gemini 4 Argon model card3. Its GraphWalks test asks a model to trace links through a huge pile of text. Up to 128,000 tokens, Argon scored 99.7 percent and GPT-6 Astra scored 98.7 percent. From 256,000 tokens to 1 million, Argon fell to 84.2 and Astra to 71.8.
Claude Opus 5.5 fell from 90.6 to 66.8 on the same test, and Claude Fable 5.1 fell from 91.4 to 65.03. Google's Gemini 4 Carbon scored 96.2 percent at the long end, the best mark on Google's page4. Google ran every score itself, and its own models lead.
Recall falls 15.5 to 26.9 points once a prompt passes 256,000 tokens on Google's long-context test
GraphWalks BFS F1 score in percent, higher is better, on the Gemini 4 Argon model card published Oct. 13, 2026
Each row is one model: the blue dot is its score on prompts up to 128,000 tokens, the magenta dot is its score from 256,000 to 1 million tokens, and the rod is the drop.
Google DeepMind ran every score on its own model card, and its own model leads, so test your own files before you trust any one gap.
Source: Google DeepMind [3]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Up to 128,000 tokens | 256,000 to 1 million tokens |
|---|---|---|
| Gemini 4 Argon | 99.7% | 84.2% |
| GPT-6 Astra | 98.7% | 71.8% |
| Claude Fable 5.1 | 91.4% | 65.0% |
| Claude Opus 5.5 | 90.6% | 66.8% |
Outside researchers found the same slide. Chroma tested 18 models in July 2025. Its report says "model performance consistently degrades with increasing input length"21. In plain words, the more you paste in, the worse the answers get, even well inside the window.
A study called NoLiMa tested 13 models that claim at least 128,000 tokens22. At just 32,000 tokens, 11 of them fell below half of their short-prompt scores. Anthropic's own engineers gave the effect a name. "As the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases," they wrote, and they called it context rot23.
The Price of a Long Prompt
Long prompts can also cost more per token. OpenAI charges GPT-6 Astra prompts over 272,000 input tokens at 2 times the input rate and 1.5 times the output rate1. That higher rate covers the full request.
Claude Haiku 5.5 input costs $0.10 per million tokens on prompts up to 100,000 tokens6. Past that line, input costs $0.50, five times as much. Meta's Muse Ember triples its input price, from $0.10 to $0.30, above 200,000 input tokens18.
What Your Chat Plan Really Gets
Most small businesses work in a chat app, where the window can be much smaller than the API's.
Claude's paid plans give the full 1 million tokens to Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.57. Claude Pro costs $20 a month36. Gemini's AI Pro plan costs $19.99 a month and also reads 1 million tokens937. A cheaper AI Plus plan holds 128,000 tokens, and free Gemini holds 32,000.
ChatGPT Plus holds 54,000 tokens in its Instant mode and 256,000 in its Reasoning mode8. Pro raises those to 128,000 and 400,000, and the free tier holds 27,000.
Claude's paid plans and Gemini AI Pro hold 1 million tokens in chat, nearly 4 times ChatGPT Plus
Context window in the chat app, in thousands of tokens, by plan and mode, from each maker's plan pages, Nov. 12, 2026
Each bar is one plan's window in the chat app; the magenta bars are the plans that give the full 1 million tokens, and gray bars are smaller chat windows.
Instant and Reasoning are ChatGPT's two chat modes, and each mode has its own window.
Sources: Claude Help Center [7]; OpenAI [8]; Gemini Apps Help [9]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Claude paid plans | 1,000K |
| Gemini AI Pro | 1,000K |
| ChatGPT Pro, Reasoning | 400K |
| ChatGPT Plus, Reasoning | 256K |
| Gemini AI Plus | 128K |
| ChatGPT Plus, Instant | 54K |
| Gemini free | 32K |
| ChatGPT Free | 27K |
Anthropic's help page says Sonnet 5 "automatically compacts the conversation at 500K tokens"7. In plain words, the app shrinks an old chat into a summary once it grows that long.
How the Pros Manage Context
Builders treat context as a budget to spend with care. They even gave the skill a name, context engineering. Tobi Lütke, who runs Shopify, called it "the art of providing all the context for the task to be plausibly solvable by the LLM"27.
Andrej Karpathy, a founding member of OpenAI, backed the term six days later. "Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down," he wrote26. In plain words, feed the model what the task needs and stop there.
Andrej Karpathy@karpathy+1 for "context engineering" over "prompt engineering".
People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window… https://t.co/Ne65F6vFcf
Start a fresh chat for each task
Simon Willison builds open source tools and writes daily about AI. "When you start a new conversation you reset that context back to zero," he wrote28. "This is important to know, as often the fix for a conversation that has stopped being useful is to wipe the slate clean and start again."
Anthropic's Claude Code guide gives the same advice for coders. It says to clear the chat "between unrelated tasks. It costs nothing, and it means the next task doesn't keep paying for the previous one's history"25.
Keep each client in its own project
Projects are folders inside a chat app that hold their own files and memory. Anthropic says Claude keeps "a separate memory for each project"35. "This ensures that your product launch planning stays separate from client work, and confidential discussions remain separate from general operations," the company wrote.
Write a handoff note
Long jobs outlast one chat, so Justin Young of Anthropic built agents that leave a progress file behind. "The key insight here was finding a way for agents to quickly understand the state of work when starting with a fresh context window," he wrote24. In plain words, end each long chat with a short note on where the work stands.
Summarize before the window fills
Anthropic's engineers call this compaction. "Compaction is the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary," they wrote23. Walden Yan, a co-founder of Cognition, built a model whose job is "to compress a history of actions & conversation into key details, events, and decisions"32.
Fetch files when you need them
Pasting every document up front fills the available window quickly. Anthropic's engineers favor a "just in time" approach, where agents keep "lightweight identifiers (file paths, stored queries, web links, etc.)" and load data only when a step calls for it23. Lance Martin of Anthropic puts it in two lines: "Use the filesystem to offload context and for progressive disclosure. Use sub-agents to isolate context"34.
Split big jobs into smaller ones
Boris Cherny created Claude Code at Anthropic. His team's tip is to "offload individual tasks to subagents to keep your main agent's context window clean and focused"30. Harrison Chase, a co-founder of LangChain, names the usual culprit when an agent fails. "Most of the time when an agent is not performing reliably the underlying cause is that the appropriate context, instructions and tools have not been communicated to the model," he wrote33.
Watch what memory keeps
Memory features carry earlier conversations into new ones, and Willison wants a say in what carries over. "I need control over what older conversations are being considered, on as fine-grained a level as possible without it being frustrating to use," he wrote29.
Drew Breunig, a writer on AI systems, explains why. "The problem is: if you put something in the context the model has to pay attention to it"31.
What to do
- Pick your chat plan by its window. Claude Pro and Gemini AI Pro each give you 1 million tokens for about $20 a month793637.
- Give each task its own conversation, and start fresh whenever the topic changes.
- Keep a one-page brief for each project and paste it at the top of every new chat.
- Put each client in a separate project.
- Keep prompts under 200,000 tokens when you can. Recall slips and some prices climb past that range3618.
- Test a long file by asking for three facts you already know.
What to watch
- Outside recall scores for Claude Fable 5.2 and GPT-6.1 Sol, once someone publishes them.
- Whether ChatGPT Plus grows its chat window toward the API's 1.05 million tokens.
Comics
Imaginary Me
A boy asks his robot to imagine a world where they never met.
Sources
- OpenAI, "GPT-6 Astra," OpenAI API documentation, Nov. 12, 2026, https://developers.openai.com/api/docs/models/gpt-6-astra
- OpenAI, "GPT-6.1 Sol," OpenAI API documentation, Nov. 12, 2026, https://developers.openai.com/api/docs/models/gpt-6.1-sol
- Google DeepMind, "Gemini 4 Argon model card," Google DeepMind, Oct. 13, 2026, https://deepmind.google/models/model-cards/gemini-4-argon/
- Google DeepMind, "Gemini models," Google DeepMind, Nov. 12, 2026, https://deepmind.google/models/gemini/
- Anthropic, "Models overview," Claude Platform documentation, Nov. 12, 2026, https://platform.claude.com/docs/en/about-claude/models/overview
- Anthropic, "Pricing," Claude Platform documentation, Nov. 12, 2026, https://platform.claude.com/docs/en/about-claude/pricing
- Anthropic, "How large is the context window on paid Claude plans?" Claude Help Center, Nov. 12, 2026, https://support.claude.com/en/articles/8606394
- OpenAI, "ChatGPT pricing," OpenAI, Nov. 12, 2026, https://chatgpt.com/pricing
- Google, "Gemini Apps limits and upgrades," Gemini Apps Help, Nov. 12, 2026, https://support.google.com/gemini/answer/16275805
- Moonshot AI, "Models," Kimi API Platform, Nov. 12, 2026, https://platform.kimi.ai/docs/models
- DeepSeek, "Models and pricing," DeepSeek API Docs, Nov. 12, 2026, https://api-docs.deepseek.com/quick_start/pricing
- Meituan, "LongCat API Platform documentation," LongCat, Nov. 12, 2026, https://longcat.chat/platform/docs
- Alibaba Cloud, "Qwen3.8-Max," Alibaba Cloud Model Studio, Nov. 5, 2026, https://www.alibabacloud.com/help/en/model-studio/qwen3-8-max
- Inception, "Models," Inception documentation, Nov. 12, 2026, https://docs.inceptionlabs.ai/get-started/models
- Z.ai, "GLM-5.3," Z.ai documentation, Nov. 12, 2026, https://docs.z.ai/guides/llm/glm-5.3
- Amazon Web Services, "What is Amazon Nova 2?" Amazon Nova User Guide, Nov. 12, 2026, https://docs.aws.amazon.com/nova/latest/nova2-userguide/what-is-nova-2.html
- xAI, "Grok 4.3," xAI documentation, Nov. 12, 2026, https://docs.x.ai/docs/models/grok-4.3
- Meta, "Meta Model API," Meta AI for Developers, Nov. 12, 2026, https://dev.meta.ai/products/meta-model-api
- MiniMax, "Models introduction," MiniMax API documentation, Nov. 12, 2026, https://platform.minimax.io/docs/guides/models-intro
- Mistral AI, "Mistral Large 4," Mistral AI documentation, Nov. 5, 2026, https://docs.mistral.ai/models/mistral-large-4
- Kelly Hong, Anton Troynikov and Jeff Huber, "Context Rot: How Increasing Input Tokens Impacts LLM Performance," Chroma, Jul. 14, 2025, https://www.trychroma.com/research/context-rot
- Ali Modarressi et al., "NoLiMa: Long-Context Evaluation Beyond Literal Matching," arXiv, Feb. 7, 2025, https://arxiv.org/abs/2502.05167
- Prithvi Rajasekaran, Ethan Dixon, Carly Ryan and Jeremy Hadfield, "Effective context engineering for AI agents," Anthropic Engineering, Sep. 29, 2025, https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
- Justin Young, "Effective harnesses for long-running agents," Anthropic Engineering, Nov. 26, 2025, https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
- Anthropic, "Best practices for Claude Code," Claude Code documentation, Nov. 12, 2026, https://code.claude.com/docs/en/best-practices
- Andrej Karpathy, "+1 for "context engineering" over "prompt engineering"." X (Andrej Karpathy), Jun. 25, 2025, https://x.com/karpathy/status/1937902205765607626
- Tobi Lütke, "I really like the term "context engineering" over prompt engineering." X (Tobi Lütke), Jun. 19, 2025, https://x.com/tobi/status/1935533422589399127
- Simon Willison, "Here's how I use LLMs to help me write code," Simon Willison's Weblog, Mar. 11, 2025, https://simonwillison.net/2025/Mar/11/using-llms-for-code/
- Simon Willison, "I really don't like ChatGPT's new memory dossier," Simon Willison's Weblog, May 21, 2025, https://simonwillison.net/2025/May/21/chatgpt-new-memory/
- Boris Cherny, "8. Use subagents," X (Boris Cherny), Jan. 31, 2026, https://x.com/bcherny/status/2017742755737555434
- Drew Breunig, "How Long Contexts Fail," dbreunig.com, Jun. 22, 2025, https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
- Walden Yan, "Don't Build Multi-Agents," Cognition, Jun. 12, 2025, https://cognition.ai/blog/dont-build-multi-agents
- Harrison Chase, "The rise of "context engineering," LangChain, Jun. 23, 2025, https://www.langchain.com/blog/the-rise-of-context-engineering
- Lance Martin, "Agent design patterns," Lance's Blog, Jan. 9, 2026, https://rlancemartin.github.io/2026/01/09/agent_design/
- Anthropic, "Bringing memory to Claude," Claude, Oct. 23, 2025, https://claude.com/resources/articles/memory
- Anthropic, "Plans and pricing," Claude, Nov. 12, 2026, https://claude.com/pricing
- Google, "Google AI plans," Gemini, Nov. 12, 2026, https://gemini.google/subscriptions/
