o3-2025-04-16· OpenAI
LMArena Human Preference 140K
list top models by context window size
- User
list top models by context window size
- Assistant
🎯 Largest-context LLMs (publicly announced, May 2024)
Rank Model Max context window (tokens*) ≈ Words Provider / Availability Key notes 1 Gemini 1.5 Pro 1 000 000 ~750 000 Google DeepMind (private preview API) “Long-context” architecture; can ingest full movies or large code bases 2 Claude 3 family (Opus / Sonnet / Haiku) 200 000 ~150 000 Anthropic (API) Same 200 k limit across tiers; streaming & tool-use ready 3 Claude 2.1 200 000 ~150 000 Anthropic (API) First mainstream 200 k release (Nov 2023) 4 Yi-34B-200K (open-source) 200 000 ~150 000 01.AI (HF download) Rotary-scaled; instruction-tuned checkpoints available 5 GPT-4 Turbo 128 000 ~95 000 OpenAI (API, ChatGPT “Turbo”) Price-tiered vs 8 k/32 k GPT-4 6 Cohere Command-R / R⁺ 128 000 ~95 000 Cohere (API) Retrieval-augmented by design 7 GPT-4 (32k variant) 32 000 ~24 000 OpenAI (legacy API tier) Original GPT-4 launch model 8 Mistral-Large 32 000 ~24 000 Mistral (API) Flash-attention 2; 27B params 9 Mixtral 8×7B Instruct (long-rope) 32 000 ~24 000 Mistral (open-source) Community “long-rope” checkpoints reach 65 k+ 10 Llama-2-70B-Long 32 000 ~24 000 Meta / community re-rope Open-source; slower inference 11 GPT-3.5-Turbo-16k 16 000 ~12 000 OpenAI (API) Budget option with extended window 12 Standard GPT-3.5-Turbo 4 096 ~3 000 OpenAI (API, free ChatGPT tier) Baseline for many apps *Token ≈ 0.75 word in English; 100 k tokens ≈ 75 k words or ~300 book pages.
Quick takeaways • If you truly need huge documents (>128 k tokens) today, Gemini 1.5 and Claude 3 are the only turnkey choices; Yi-34B-200K is an open-source DIY option.
• For most enterprise / RAG workloads, 32 k–128 k windows (GPT-4 Turbo, Cohere Command-R, Mistral-Large) already cover entire legal contracts, codebases, or multi-chapter manuals.
• Very long windows do not guarantee perfect recall—chunking + retrieval still matters! 🧠🔍
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.