📏

AI Model Context Window Cheat Sheet

Compare input context limits and max output tokens across frontier, balanced, fast, reasoning, and open-weight LLMs.

Filter by category

Provider Model Context Max output Category Input / 1M Output / 1M
OpenAI OpenAI GPT-4.1

1M context, strong long-context retrieval.

1.0M 33K Frontier $2.00 $8.00
OpenAI OpenAI GPT-4.1 mini

Same context window as GPT-4.1 at much lower cost.

1.0M 33K Balanced $0.40 $1.60
OpenAI OpenAI GPT-4.1 nano

Smallest, cheapest, fastest; still 1M context.

1.0M 33K Fast / cheap $0.10 $0.40
OpenAI OpenAI GPT-4o

128K context, natively multimodal.

128K 16K Frontier $2.50 $10.00
OpenAI OpenAI o3-mini

Reasoning model; high output budget, 200K context.

200K 100K Reasoning $1.10 $4.40
OpenAI OpenAI o1

Deep reasoning; expensive but powerful for hard STEM.

200K 100K Reasoning $15.00 $60.00
Anthropic Anthropic Claude 4 Sonnet

200K context; strong agentic coding and analysis.

200K 8K Frontier $3.00 $15.00
Anthropic Anthropic Claude 4 Opus

Highest-end Claude; 200K context.

200K 8K Frontier $15.00 $75.00
Anthropic Anthropic Claude 4 Haiku

Fast and cheap; same 200K window.

200K 8K Fast / cheap $0.25 $1.25
Google Google Gemini 2.5 Pro

1M token context, very large output budget.

1.0M 66K Frontier $1.25 $10.00
Google Google Gemini 2.5 Flash

Fast, cheap, 1M context.

1.0M 66K Balanced $0.15 $0.60
Google Google Gemini 2.0 Flash

1M context at very low cost.

1.0M 8K Fast / cheap $0.10 $0.40
DeepSeek DeepSeek-V3

Low API cost; 64K context.

64K 8K Balanced $0.07 $1.10
DeepSeek DeepSeek-R1

Open reasoning model; 64K context.

64K 8K Reasoning $0.14 $2.19
Meta (open weights) Meta Llama 4 Scout

Open weights; 128K context, can run locally.

131K 8K Open weights
Meta (open weights) Meta Llama 4 Maverick

Open weights; larger variant, 128K context.

131K 8K Open weights
Alibaba Cloud Alibaba Qwen3-235B-A22B

Mixture-of-experts; 128K context.

128K 8K Frontier $0.50 $2.00
Alibaba Cloud Alibaba Qwen3-32B

Strong smaller model; 128K context.

128K 8K Balanced $0.10 $0.30
Mistral AI Mistral Large 2

128K context, strong multilingual and coding.

128K 8K Frontier $2.00 $6.00
Cohere Cohere Command R+

128K context; built for RAG and enterprise.

128K 8K Balanced $2.50 $10.00
xAI xAI Grok 3

128K context; reasoning and coding focused.

128K 8K Frontier $5.00 $25.00

Key rules of thumb

  • System prompt + tools + RAG + history must fit inside the context window. Reserve ~10-20% as a safety buffer.
  • Max output tokens limits how long the model's response can be. Reasoning models often have larger output budgets.
  • 1M-token contexts (OpenAI GPT-4.1, Google Gemini 2.5) fit entire codebases, long documents, or multi-turn sessions.
  • 200K contexts (Claude 4) still cover most enterprise use cases with strong retrieval quality.
  • Open-weight models (Llama 4, Qwen 3) are often limited by local GPU VRAM rather than the model's theoretical context.

Token budget planner

Total required context

Frequently asked questions

What is a context window in an LLM?

The context window is the maximum number of input tokens a model can process in one request. It includes the system prompt, user messages, few-shot examples, retrieved RAG chunks, tool outputs, and conversation history.

What is the difference between context tokens and max output tokens?

Context tokens are the total tokens the model can read. Max output tokens are the tokens it is allowed to generate in its response. Both consume inference time and cost, but providers usually price them separately.

Which model has the largest context window?

As of mid-2026, OpenAI GPT-4.1 series and Google Gemini 2.5 series support roughly 1 million context tokens. Anthropic Claude 4 series supports 200K tokens. Always check the provider's latest docs.

Do larger context windows always mean better retrieval?

Not necessarily. A model can read a long context, but retrieval quality, needle-in-haystack performance, and cost vary. Long-context models work best when the relevant information is clearly formatted and placed.

How do I choose a context window for my agent?

Estimate your system prompt + tool definitions + RAG chunks + user history + desired output reserve. Add a 20% safety buffer. If you plan to ingest full documents or long conversations, prioritize 128K+ models.

Last updated: 2026-07-28. See notes.

Context-window and output-token limits are based on public API documentation. Providers change limits, pricing, and model availability frequently; verify before building production agents.

🚀 Get AI automation insights daily

15:00 MST. One-click unsubscribe.

Subscribe