Context window comparison: every major AI model in one table

The quick answer

The largest context window in the current catalogue is 1M tokens, held by DeepSeek V4 Flash 0731 from DeepSeek. As a rough physical intuition, one token is about three quarters of an English word, and a manuscript page is about 400 words, so the table below translates every window into pages.

One row per major provider, showing its largest-context model:

ProviderLargest-context modelContext windowFits roughly
DeepSeekDeepSeek V4 Flash 07311M tokensabout 2,500 pages
OpenAIGPT-5.41M tokensabout 2,000 pages
GoogleGemini 2.5 Flash1M tokensabout 2,000 pages
MetaLlama 4 Maverick1M tokensabout 2,000 pages
QwenQwen3.8 2.4T A95B1M tokensabout 2,000 pages
MoonshotKimi K31M tokensabout 2,000 pages
Z.aiGLM 5.21M tokensabout 2,000 pages
MiniMaxMiniMax M31M tokensabout 2,000 pages
AnthropicClaude Fable 51M tokensabout 1,900 pages
xAIGrok 4.31M tokensabout 1,900 pages
MistralMistral Large 3 2512262K tokensabout 492 pages

If the question behind your search is "what is a context window in the first place", start with the plain-English explainer at What is a context window? and come back for the numbers.

Every model in the index (100 models)

The full catalogue, sorted by context window. The data comes from the same source as the AI Model Cost Index and was last fetched 2026-08-20; when the catalogue updates, this table updates with it. The credits column shows what each model costs per message inside Whizi, where a plan includes it.

ModelProviderContext windowFits roughlyCredits in Whizi
DeepSeek V4 Flash 0731DeepSeek1Mabout 2,500 pages1
GPT-5.4OpenAI1Mabout 2,000 pages15
GPT-5.5OpenAI1Mabout 2,000 pages20
GPT-5.6 LunaOpenAI1Mabout 2,000 pages1
GPT-5.6 Luna ProOpenAI1Mabout 2,000 pagesNot metered
GPT-5.6 SolOpenAI1Mabout 2,000 pages20
GPT-5.6 Sol ProOpenAI1Mabout 2,000 pagesNot metered
GPT-5.6 TerraOpenAI1Mabout 2,000 pages4
GPT-5.6 Terra ProOpenAI1Mabout 2,000 pagesNot metered
LongCat 2.0Meituan1Mabout 2,000 pages1
DeepSeek V4 Pro 0813DeepSeek1Mabout 2,000 pages2
Gemini 2.5 FlashGoogle1Mabout 2,000 pages2
Gemini 2.5 Flash LiteGoogle1Mabout 2,000 pages1
Gemini 3.1 Flash LiteGoogle1Mabout 2,000 pages1
Gemini 3.1 Pro PreviewGoogle1Mabout 2,000 pages10
Gemini 3.5 FlashGoogle1Mabout 2,000 pages8
Gemini 3.5 Flash LiteGoogle1Mabout 2,000 pages2
Gemini 3.6 FlashGoogle1Mabout 2,000 pages3
Gemini 3.7 FlashGoogle1Mabout 2,000 pages2
GLM 5.2Z.ai1Mabout 2,000 pages3
GLM 5.3Z.ai1Mabout 2,000 pages5
InklingThinkingmachines1Mabout 2,000 pages4
Kimi K3Moonshot1Mabout 2,000 pages10
Laguna S 2.1Poolside1Mabout 2,000 pages1
Llama 4 MaverickMeta1Mabout 2,000 pages1
MiniMax M3MiniMax1Mabout 2,000 pages1
Muse Spark 1.1Meta1Mabout 2,000 pages5
Muse Spark 1.2Meta1Mabout 2,000 pages5
Qwen3.8 2.4T A95BQwen1Mabout 2,000 pages8
Claude Fable 5Anthropic1Mabout 1,900 pages50
Claude Opus 4.7 (Fast)Anthropic1Mabout 1,900 pagesNot metered
Claude Opus 4.8Anthropic1Mabout 1,900 pages20
Claude Opus 4.8 (Fast)Anthropic1Mabout 1,900 pagesNot metered
Claude Opus 5Anthropic1Mabout 1,900 pages20
Claude Opus 5 (Fast)Anthropic1Mabout 1,900 pagesNot metered
Claude Sonnet 4.5Anthropic1Mabout 1,900 pages10
Claude Sonnet 4.6Anthropic1Mabout 1,900 pages10
Claude Sonnet 5Anthropic1Mabout 1,900 pages10
Fugu UltraSakana1Mabout 1,900 pages25
Grok 4.3xAI1Mabout 1,900 pages4
Qwen3.5 Plus 2026-04-20Qwen1Mabout 1,900 pages1
Qwen3.6 FlashQwen1Mabout 1,900 pages1
Qwen3.7 FlashQwen1Mabout 1,900 pages1
Qwen3.7 MaxQwen1Mabout 1,900 pages5
Qwen3.7 PlusQwen1Mabout 1,900 pages1
Qwen3.8 27BQwen1Mabout 1,900 pages3
Qwen3.8 MaxQwen1Mabout 1,900 pages6
Inkling SmallThinkingmachines524Kabout 983 pages1
Solar Pro 4Upstage524Kabout 983 pages1
Nemotron 3 UltraNVIDIA512Kabout 961 pages3
Grok 4.5xAI500Kabout 938 pages6
Grok 4.6xAI500Kabout 938 pages6
GPT Chat LatestOpenAI400Kabout 750 pages25
GPT-5.4 MiniOpenAI400Kabout 750 pages4
GPT-5.4 NanoOpenAI400Kabout 750 pages2
Nova Pro 1.0Amazon300Kabout 563 pages3
Hy3Tencent262Kabout 492 pages1
Kimi K2 ThinkingMoonshot262Kabout 492 pages10
Kimi K2.7 CodeMoonshot262Kabout 492 pages10
Laguna XS 2.1Poolside262Kabout 492 pages1
Ling-3.0-flashinclusionAI262Kabout 492 pages1
Mistral Large 3 2512Mistral262Kabout 492 pages2
Mistral Medium 3.5Mistral262Kabout 492 pages6
Nemotron 3.5 LightningNVIDIA262Kabout 492 pages1
Nex-N2-MiniNex Agi262Kabout 492 pages1
Nex-N2-ProNex Agi262Kabout 492 pages1
Qwen3 Coder NextQwen262Kabout 492 pages1
Qwen3.6 35B A3BQwen262Kabout 492 pages1
Ring-2.6-1TinclusionAI262Kabout 492 pages1
Sakana NamazuSakana262Kabout 492 pages4
Seed 2.1 TurboBytedance Seed262Kabout 492 pages2
Seed-2.0-CodeBytedance Seed262Kabout 492 pages3
Step 3.7 FlashStepfun262Kabout 492 pages1
Codestral 2508Mistral256Kabout 480 pages1
Command ACohere256Kabout 480 pages10
Grok Build 0.1xAI256Kabout 480 pages3
KAT-Coder-Air V2.5Kwaipilot256Kabout 480 pages1
KAT-Coder-Pro V2.5Kwaipilot256Kabout 480 pages3
GLM 4.6Z.ai205Kabout 384 pages1
GLM 4.7Z.ai205Kabout 384 pages2
GLM 5Z.ai205Kabout 384 pages2
MiniMax M2MiniMax205Kabout 384 pages1
Claude Haiku 4.5Anthropic200Kabout 375 pages4
DeepSeek V3.1DeepSeek164Kabout 307 pages1
DeepSeek V3.2DeepSeek164Kabout 307 pages1
Aion-3.0Aion Labs131Kabout 246 pages10
Aion-3.0-MiniAion Labs131Kabout 246 pages2
Granite 4.1 8BIbm Granite131Kabout 246 pages1
Kimi K2 0711Moonshot131Kabout 246 pages10
Llama 3.3 70B InstructMeta131Kabout 246 pages1
Mistral Medium 3.1Mistral131Kabout 246 pages2
Muse Glimmer 30BMeta131Kabout 246 pages2
Nano Banana 2 (Gemini 3.1 Flash Image)Google131Kabout 246 pagesNot metered
Nano Banana Pro (Gemini 3 Pro Image)Google131Kabout 246 pagesNot metered
Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image)Google66Kabout 123 pagesNot metered
R1DeepSeek64Kabout 120 pages2
Perceptron Mk1Perceptron33Kabout 61 pages1
Phi 4Microsoft16Kabout 31 pages1
Hy-MT2-1.8BTencent8Kabout 15 pagesNot metered
Hy-MT2-30B-A3BTencent8Kabout 15 pagesNot metered

How to read these numbers honestly

The window is shared. Everything counts against it at once: your prompt, every earlier message in the conversation, any documents you attached, and the answer being written. A model with a 128K window that has already read a 100K-token contract does not have 128K left for the discussion; it has 28K.

Advertised is not effective. Long-context research keeps finding the same pattern: models recall the start and end of a huge context better than the middle. A model can accept a million tokens and still gloss over the clause buried at token 400,000. For high-stakes work on long documents, ask specifically about the middle sections, or feed the document in labelled parts.

Bigger is not automatically better. Large windows cost more to run and can dilute attention across irrelevant text. If your work is emails and short drafts, the window size will never be the thing you notice. It starts to matter at whole contracts, codebases, transcripts, and books.

Tokens are not words. The three-quarters rule is a good English average, but code, numbers, and non-English languages often tokenize less efficiently, so the page estimates in the table are deliberately rough.

Using the right window for the job

The practical conclusion from the table is that context window is a per-task choice, not a per-subscription choice. The model with the biggest window is rarely the best writer, and the best writer is rarely the cheapest way to summarize 800 pages.

That is the argument for having the catalogue in one workspace. In Whizi you can hand the 200-page PDF to a large-context model, then switch to Claude or GPT in the same conversation to draft from what it found, without the document leaving the thread. The how to choose a model guide covers the same decision for quality and cost.

For citing this page: the anchor-free URL is stable, the data date is printed above the full table, and the token figures come from the provider catalogue rather than from marketing pages.

Workflow checklist
  • Subtract your documents and history from the window before trusting the headline number
  • For long documents, test recall from the middle, not just the start
  • Translate tokens to pages with the three-quarters rule before deciding anything
  • Pick the window per task instead of per subscription
  • Check the data date above the full table when citing a figure
Common questions

Frequently asked questions

Which AI model has the largest context window?

In the current catalogue it is DeepSeek V4 Flash 0731 from DeepSeek at 1M tokens. The full table above lists all 100 models sorted by window size, with the data date printed above it.

What is the largest Claude context window?

The largest Claude context window in the index is 1M tokens (Claude Fable 5), which fits about 1,900 pages of text. Smaller Claude variants are listed in the full table above.

What is the largest GPT context window?

The largest GPT context window in the index is 1M tokens (GPT-5.4), which fits about 2,000 pages of text. Smaller GPT variants are listed in the full table above.

What is the largest Gemini context window?

The largest Gemini context window in the index is 1M tokens (Gemini 2.5 Flash), which fits about 2,000 pages of text. Smaller Gemini variants are listed in the full table above.

What is the largest Grok context window?

The largest Grok context window in the index is 1M tokens (Grok 4.3), which fits about 1,900 pages of text. Smaller Grok variants are listed in the full table above.

Is a bigger context window always better?

No. Larger windows cost more, and recall degrades toward the middle of a very large context. Bigger is decisive for whole books, codebases, and long transcripts, and irrelevant for everyday chat. Match the window to the task rather than paying for the biggest number.

How many pages fit in a context window?

Divide the token count by roughly 530 to get manuscript pages: one token is about three quarters of an English word, and a page is about 400 words. So 128K tokens is about 240 pages, and a million tokens is about 1,900 pages. Code and non-English text usually fit less.