LLM Context Window Comparison: Largest Models (2026)

Which LLM has the largest context window? As of October 2026, Meta’s open-weight Llama 4 Scout advertises the most, at 10M tokens. Among hosted APIs, the ceiling is about 1M tokens: OpenAI’s GPT-6 family lists 1.05M, and current Claude, Gemini 3.x, DeepSeek V4 and Grok 4.3 models list 1M. Max output ranges from 64K to 384K.

The rest of this page is a sourced LLM context window comparison: hosted models, open-weight models, what long prompts cost, and why the biggest number isn’t always the one to pick. If you want the basics first, read what the context window of an LLM is.

Which LLM has the largest context window?

The largest context window in October 2026 belongs to Llama 4 Scout at 10M tokens, followed by a large group of models at about 1M tokens. The group at 1M includes the flagship models from OpenAI, Anthropic, Google, DeepSeek and xAI, plus several open-weight models such as Kimi K3, GLM-5.3 and MiniMax-M3.

Two cautions apply to any context window ranking:

• Advertised is not effective. The listed size is the most the API accepts, not the most the model reads well. More on this below.
• Tokens aren’t equal across models. Each provider has its own tokenizer. Anthropic says 1M tokens is about 555K words on its current tokenizer, versus about 750K words on older Claude models.

Context window comparison table: hosted APIs

Every number below comes from the provider’s own documentation, checked on October 8, 2026.

As of October 2026. Model lineups change every few months; confirm on the provider page before you depend on a number.

A few things stand out:

• 1M is the new default. Anthropic says 1M is the standard window on all its current models, with no beta header needed. Older Claude models, such as Claude Sonnet 4.5, stay at 200K.
• OpenAI’s window includes output. GPT-6.1 Sol’s 1.05M window allows up to 922,000 input tokens, leaving room for its 128,000-token output.
• DeepSeek allows the longest output in this list, at up to 384K tokens.
• xAI’s newest model has a smaller window. grok-4.7 lists 500K, half of the older grok-4.3.

Open-weight models with the largest context windows

Open-weight models publish their limits in the model card. These numbers come from the official Hugging Face cards and config.json files.

As of October 2026.

Running these at full length takes a lot of GPU memory, mostly for the KV cache. The details, plus Ollama and llama.cpp settings, are in our guide to the largest context window open-source LLMs.

Long-context pricing: what a full window costs

Some providers charge a higher rate once a prompt passes a size threshold. This changes the real cost of “using the whole window” more than the headline price does.

As of October 2026. Sources: OpenAI, Gemini pricing, Anthropic, xAI, DeepSeek.

A worked example: a single 500K-token prompt to GPT-6.1 Sol crosses the 272K line, so input costs $4 per million instead of $2. That’s $2.00 of input for one call. An agent that resends that context on 20 turns spends $40 on input alone, before output. Prompt caching cuts this a lot, but cached tokens still fill the window.

Bigger isn’t always better: advertised vs effective context

A context window ranking by size says nothing about how well each model reads at that length. Independent benchmarks keep finding a gap:

• RULER tested 17 models that all claimed at least 32K tokens; only half performed well at 32K (Hsieh et al., 2024).
• NoLiMa tested 13 models that claim at least 128K tokens; at 32K, 11 of them fell below 50% of their short-context scores (Modarressi et al., 2025).

Those studies predate the current 1M-token models, and newer models handle length better. Anthropic’s own docs still warn that “as token count grows, accuracy and recall degrade.” The research and the fixes are covered in context window limitations and solutions.

The practical takeaway: pick a model by how it performs at the length you’ll actually use, not by the ceiling.

How to check a model’s context window yourself

Lists like this one go stale. Here’s where to find the current number:

1. Anthropic: the Models API returns max_input_tokens and max_tokens for every model.
2. OpenAI: each model page lists context window, max input and max output.
3. Google: each Gemini model page lists the input and output token limits.
4. DeepSeek and xAI: the pricing pages list context length next to price.
5. Open-weight models: read the model card, then check max_position_embeddings in config.json.

For open models, a short script reads the config straight from Hugging Face:

 1import json
 2import urllib.request
 3
 4def context_length(repo: str):
 5    url = f"https://huggingface.co/{repo}/raw/main/config.json"
 6    with urllib.request.urlopen(url) as resp:
 7        cfg = json.load(resp)
 8    cfg = cfg.get("text_config", cfg)  # multimodal models nest the text settings
 9    return cfg.get("max_position_embeddings"), cfg.get("rope_scaling")
10
11for repo in ["Qwen/Qwen3.6-27B", "openai/gpt-oss-120b", "zai-org/GLM-5.3"]:
12    print(repo, context_length(repo))
13# Qwen/Qwen3.6-27B (262144, None)
14# openai/gpt-oss-120b (131072, {... 'rope_type': 'yarn' ...})
15# zai-org/GLM-5.3 (1048576, None)

Treat the config value as a hint and the model card as the authority. Qwen3-8B’s config says 40,960, while its card says 32,768 natively and 131,072 with YaRN. Gated models such as Llama 4 require a Hugging Face login to read the config.

Choosing a model by context needs

• Under 128K tokens (most chat and RAG apps): almost any current model fits. Choose on quality and price.
• 128K to 1M tokens (whole codebases, long contracts, long agent runs): any current Claude, GPT-6, Gemini 3.x, DeepSeek V4 or Grok 4.3 model. Check the long-context price tier.
• Over 1M tokens: only Llama 4 Scout advertises it, and you host it yourself. Usually it’s cheaper and more accurate to retrieve the relevant parts instead; see techniques for extending an LLM’s context window.
• Running locally: the limit is your GPU memory, not the model card.
• Agents whose history keeps growing: no window holds months of sessions. An external memory layer such as Hindsight (MIT) or Mem0 stores facts outside the model and sends back only the relevant ones; Hindsight posts the top vendor-reported scores on the BEAM 1M and 10M-token history tests. See Mem0 vs Hindsight.