Who ranks #1 on the Long Context leaderboard?
As of August 21, 2026, Gemini 3.7 Flash by Google ranks #1 on MRCR v2 at 97%. API pricing is $0.75/M input and $3.75/M output.
Long Context
🏆 LeaderboardAs of August 21, 2026, Gemini 3.7 Flash is #1 for long-context retrieval at 97%. Ranked by MRCR: can the model still find planted facts inside a long prompt. Claude vs Gemini context window, ranked by MRCR and Graphwalks. A 1M context window is not the same as usable long context.
Updated August 21, 2026282 models33 providers
Showing 1–50 of 282 models
Next step
You found the model. Now ship the product.
Auth, billing, and the API layer are already decided. Fork 8 finished AI apps, or have us build the first version with you.
The index
All Large Language Models
282 models across 33 providers. Search or jump to a lab — every model page stays linked here.
OpenAI
51 modelsGPT-5.6 CyberGPT-5.6 LunaGPT-5.6 SolGPT-5.6 TerraGPT-5.5 InstantGPT-5.5GPT-5.5 ProGPT-5.4 miniGPT-5.4 nanoGPT-5.4GPT-5.3 ChatGPT-5.3 CodexGPT-5.2 CodexGPT-5.2GPT-5.2 ProGPT-5.1 CodexGPT-5.1GPT-5.1 Codex HighGPT-5.1 Codex MiniGPT-5.1 HighGPT-5.1 InstantGPT-5.1 MediumGPT-5.1 ThinkingGPT-5 CodexGPT-5GPT-5 HighGPT-5 MediumGPT-5 miniGPT-5 nanoGPT OSS 120BGPT OSS 120B HighGPT OSS 20BGPT OSS 20B Higho3-proo3o4-miniGPT-4.1GPT-4.1 miniGPT-4.1 nanoo3-minio1o1-proo1-minio1-previewGPT-4oGPT-4o miniChatGPT-4o LatestGPT-4oGPT-4 TurboGPT-4GPT-3.5 Turbo
Anthropic
22 modelsClaude Opus 5Claude Sonnet 5Claude Fable 5Claude Mythos 5Claude Opus 4.8Claude Opus 4.7Claude Sonnet 4.6Claude Opus 4.6Claude Opus 4.5Claude Haiku 4.5Claude Sonnet 4.5Claude Opus 4.1Claude Opus 4Claude Sonnet 4Claude 3.7 SonnetClaude 3.5 HaikuClaude 3.5 SonnetClaude 3.5 SonnetClaude 3 HaikuClaude 3 OpusClaude 3 SonnetClaude Mythos Preview
Gemini 3.7 FlashGemini 3.5 Flash-LiteGemini 3.6 FlashGemini 3.5 FlashGemma 4 26B-A4BGemma 4 31BGemini 3.1 Flash-LiteGemini 3.1 ProGemini 3 FlashGemini 3 ProGemma 3n E4B InstructedGemini 2.5 Flash-LiteGemini 2.5 Pro Preview 06-05Gemini 2.5 FlashGemini 2.5 ProGemma 3 12BGemma 3 27BGemma 3 4BGemini 2.0 Flash-LiteGemini 2.0 FlashGemma 2 27BGemini 1.5 FlashGemini 1.5 ProGemini 1.5 Flash 8BGemini 1.0 Pro
Cursor
2 models
xAI
15 models
DeepSeek
16 models
Qwen
38 modelsQwen3.8-27BQwen3.8-2.4T-A95BQwen3.8 MaxQwen3.7 FlashQwen3.7-PlusQwen3.7 MaxQwen3.6-27BQwen3.6-35B-A3BQwen3.6 PlusQwen3.5-9BQwen3.5-122B-A10BQwen3.5-27BQwen3.5-35B-A3BQwen3.5-397B-A17BQwen3 MaxQwen3 VL 235B A22B InstructQwen3 VL 235B A22B ThinkingQwen3 VL 30B A3B InstructQwen3 VL 30B A3B ThinkingQwen3 VL 32B InstructQwen3 VL 4B InstructQwen3 VL 4B ThinkingQwen3 VL 8B InstructQwen3 VL 8B ThinkingQwen3-Next-80B-A3B-InstructQwen3-Next-80B-A3B-ThinkingQwen3-235B-A22B-Thinking-2507Qwen3-235B-A22B-Instruct-2507Qwen3 235B A22BQwen3 30B A3BQwen3 32BQwen3-Coder 480B A35B InstructQwen2.5 VL 72B InstructQwen3-CoderQwQ-32B-PreviewQwen2.5 72B InstructQwen2.5 7B InstructQwen2.5-Coder 32B Instruct
AI21 Labs
2 models
Amazon
5 models
Baidu
1 models
ByteDance
2 models
Cohere
3 models
IBM
1 models
Inception
1 models
inclusionAI
1 models
LG AI Research
1 models
Meituan
4 models
Meta
12 models
Mistral
23 modelsMistral Medium 3.5Mistral Small 4Min istral 3 (3B Reasoning 2512)MiniStral 3 (14B Instruct 2512)Ministral 3 (14B Reasoning 2512)Ministral 3 (3B Instruct 2512)Ministral 3 (8B Instruct 2512)Ministral 3 (8B Reasoning 2512)Mistral Large 3 (675B Instruct 2512)Mistral Large 3Devstral Small 1.1Devstral MediumMistral Small 3.2 24B InstructMagistral MediumMistral Small 3.1 24B BaseMistral Small 3.1 24B InstructMistral Small 3 24B InstructPixtral LargeMinistral 8B InstructMistral SmallPixtral-12BMistral Large 2Mistral NeMo Instruct
Moonshot AI
9 models
Nous Research
1 models
NVIDIA
5 models
Poolside
2 models
Sakana AI
1 models
StepFun
1 models
Tencent
1 models
Thinking Machines
1 models
Unisound
1 models
Upstage
1 models
Xiaomi
3 modelsDoes a 1 million token context window actually work?
As of August 21, 2026, Gemini 3.7 Flash by Google is #1 for long-context retrieval at 97%. Ranked by MRCR: can the model still find planted facts inside a long prompt. This board also tracks MRCR v2, Fiction.liveBench, AA-LCR, Graphwalks BFS (0K-128K). Next on the same board: Qwen3.8 Max and Qwen3.7-Plus. Related leaders: o3 on Fiction.liveBench at 100%; Muse Glimmer-30B on AA-LCR at 80%. This long context leaderboard ranks models by MRCR v2. Scores come from public evals. Prices are the live API rates in the table above.
Sources: OpenAI API pricing (https://developers.openai.com/api/docs/pricing); Anthropic Claude API pricing (https://platform.claude.com/docs/en/about-claude/pricing); Gemini API pricing (https://ai.google.dev/gemini-api/docs/pricing)
| Rank | Model | MRCR v2 | Input /M | Output /M |
|---|---|---|---|---|
| 1 | Gemini 3.7 Flash | 97% | $0.75 | $3.75 |
| 2 | Qwen3.8 Max | 92.9% | $2.00 | $6.00 |
| 3 | Qwen3.7-Plus | 91.7% | $0.32 | $1.28 |
| 4 | GPT-5.6 Sol | 91.5% | $5.00 | $30.00 |
| 5 | GPT-5.6 Terra | 89.6% | $2.00 | $12.00 |
| 6 | Claude Opus 4.6 | 76% | $5.00 | $25.00 |
| 7 | GPT-5.5 | 74% | $5.00 | $30.00 |
| 8 | Gemma 4 31B | 66.4% | $0.14 | $0.40 |
Claude vs Gemini context window
People search Claude context window and Gemini context window more than best long context LLM. The listed window is the cap. MRCR and Graphwalks are whether the model can still find needles at 128K and beyond. A 1M window that forgets after 64K is a billing trap.
What is needle in a haystack for LLMs?
Needle-in-a-haystack plants a fact in a long prompt and asks for it back. MRCR is the harder multi-needle version. That is the eval to sort if you were burned by a model that “supports 1M” and still misses the clause on page 40.
Related LLM leaderboards
Individual benchmark pages
Rank one eval at a time. All LLM benchmarks.
Long Context FAQ
Does a 1 million token context window actually work?
As of August 21, 2026, Gemini 3.7 Flash by Google is #1 for long-context retrieval at 97%. Ranked by MRCR: can the model still find planted facts inside a long prompt. This board also tracks MRCR v2, Fiction.liveBench, AA-LCR, Graphwalks BFS (0K-128K). Next on the same board: Qwen3.8 Max and Qwen3.7-Plus. Related leaders: o3 on Fiction.liveBench at 100%; Muse Glimmer-30B on AA-LCR at 80%.
What are the top models on MRCR v2?
The current MRCR v2 ranking as of August 21, 2026 is 1. Gemini 3.7 Flash at 97%; 2. Qwen3.8 Max at 92.9%; 3. Qwen3.7-Plus at 91.7%.
Which long context model is the cheapest?
Gemma 3 27B is the cheapest scored model on this long context leaderboard at $0.10/M input and $0.20/M output ($0.30 blended). Gemini 3.7 Flash still leads MRCR v2 at 97%.
Should I always pick the #1 MRCR v2 model?
Not automatically. Gemini 3.7 Flash leads MRCR v2, but a cheaper scored model can be the better production choice if the quality gap is small. Use the table to weigh MRCR v2 against input/output price, context window, and related evals.
How often is the Long Context leaderboard updated?
Scores and API prices on this page are refreshed from published evals and provider rates. The snapshot is labeled August 21, 2026. Treat it as a current index, not a one-off blog post.
What is the best AI for long context?
A 1M context window is not enough. MRCR and Graphwalks measure whether the model can still retrieve and reason inside that window.
What is MRCR v2?
MRCR v2 is OpenAI's multi-needle retrieval eval. It hides several facts in long prompts and checks whether the model can pull them back.
Should I just pick the largest context window?
No. Some 1M models lose accuracy after 128K. Sort by MRCR or Graphwalks, then check the listed context window and token price.
What is Claude's context window?
Claude's listed context window is on each model row. Listed size is not usable size. MRCR and Graphwalks measure whether the model still retrieves facts at 128K, 256K, and 1M. Sort those columns, then read the sticker context number.
What is Gemini's context window?
Gemini advertises very large context, including 1M-class windows. The question people mean is whether retrieval still works at that length. This page ranks MRCR and Graphwalks, not the marketing number.
What is MRCR?
MRCR (Multi-Round Coreference / multi-needle retrieval) hides several facts in a long prompt and checks whether the model can pull them back. It is the usual public long-context exam next to needle-in-a-haystack demos.
Should I use RAG instead of a huge context window?
Usually yes for production. Filling 1M tokens is slow and expensive. Use long context when the document must stay in-prompt. Use RAG when you can retrieve. Sort MRCR to see if the giant window is real.