Glossary term
LMSYS Chatbot Arena Leaderboard
What is the LMSYS Chatbot Arena Leaderboard?
The LMSYS Chatbot Arena Leaderboard is a comprehensive ranking platform that assesses the performance of large language models (LLMs) in conversational tasks. It uses a combination of human feedback and automated scoring to evaluate models like GPT-4, Claude, and others, providing a clear view of their strengths and weaknesses in real-world applications.
The leaderboard is updated regularly, reflecting the latest advancements in AI technology. It includes models from leading organizations such as OpenAI, Anthropic, Google, and Meta, showcasing their capabilities in engaging and informative conversations.
Key features of the LMSYS Chatbot Arena Leaderboard include:
- Diverse Evaluation Metrics — The leaderboard uses multiple metrics for evaluation, including Chatbot Arena Elo, which is based on user feedback, and other performance indicators.
- Regular Updates — The leaderboard is frequently updated to reflect the latest developments in LLM technology, ensuring that it remains a relevant and valuable resource for AI researchers and developers.
- Community Engagement — Users can participate in the evaluation process by providing feedback on chatbot interactions, contributing to the dynamic nature of the leaderboard.
LMSYS Chatbot Arena Leaderboard (September 2024)
Historical table retained with its original breadth. Its Arena, MT-Bench, and MMLU columns have incompatible units and imperfect provenance; they are preserved as published and must not be combined.
| Model | Arena Elo rating | MT-bench (score) | MMLU | License |
|---|---|---|---|---|
| o1-preview | 1355 | +12/-11 | 2991 | Proprietary |
| ChatGPT-4o-latest (2024-09-03) | 1335 | +5/-6 | 10213 | Proprietary |
| o1-mini | 1324 | +12/-9 | 3009 | Proprietary |
| Gemini-1.5-Pro-Exp-0827 | 1299 | +5/-4 | 28229 | Proprietary |
| Grok-2-08-13 | 1294 | +4/-4 | 23999 | Proprietary |
| GPT-4o-2024-05-13 | 1285 | +3/-3 | 90695 | Proprietary |
| GPT-4o-mini-2024-07-18 | 1273 | +3/-3 | 30434 | Proprietary |
| Claude 3.5 Sonnet | 1269 | +3/-3 | 62977 | Proprietary |
| Gemini-1.5-Flash-Exp-0827 | 1269 | +4/-4 | 22264 | Proprietary |
| Grok-2-Mini-08-13 | 1267 | +4/-5 | 22041 | Proprietary |
| Gemini Advanced App (2024-05-14) | 1267 | +3/-3 | 52218 | Proprietary |
| Meta-Llama-3.1-405b-Instruct-fp8 | 1266 | +4/-4 | 31280 | Llama 3.1 Community |
| Meta-Llama-3.1-405b-Instruct-bf16 | 1264 | +6/-8 | 5865 | Llama 3.1 Community |
| GPT-4o-2024-08-06 | 1263 | +4/-3 | 22562 | Proprietary |
| Gemini-1.5-Pro-001 | 1259 | +3/-3 | 80656 | Proprietary |
| GPT-4-Turbo-2024-04-09 | 1257 | +3/-2 | 92973 | Proprietary |
| GPT-4-1106-preview | 1251 | 9.40 | 2023/4 | Proprietary |
| Mistral-Large-2407 | 1250 | — | 2024/7 | Mistral Research |
| Athene-70b | 1250 | — | 2024/7 | CC-BY-NC-4.0 |
| Meta-Llama-3.1-70b-Instruct | 1249 | — | 2023/12 | Llama 3.1 Community |
| Claude 3 Opus | 1248 | 9.45 | 87.1 | Proprietary |
| GPT-4-0125-preview | 1245 | 9.38 | — | Proprietary |
| Yi-Large-preview | 1240 | — | — | Proprietary |
| Gemini-1.5-Flash-001 | 1227 | — | 78.9 | Proprietary |
| Deepseek-v2-API-0628 | 1219 | — | — | DeepSeek |
| Gemma-2-27b-it | 1218 | — | — | Gemma license |
| Yi-Large | 1212 | — | — | Proprietary |
| Gemini App (2024-01-24) | 1209 | — | — | Proprietary |
| Nemotron-4-340B-Instruct | 1209 | — | — | NVIDIA Open Model |
| GLM-4-0520 | 1207 | — | — | Proprietary |
| Llama-3-70b-Instruct | 1206 | — | 82.0 | Llama 3 Community |
| Claude 3 Sonnet | 1201 | 9.22 | 87.0 | Proprietary |
Vision Leaderboard
The Vision Leaderboard ranks top LLMs based on their performance in vision-based conversations. It evaluates models using metrics like Arena Elo, MT-bench score, and MMLU, providing insights into their strengths and weaknesses.
| Model | Arena Score | Organization | Knowledge Cutoff |
|---|---|---|---|
| Gemini-1.5-Pro-Exp-0827 | 1231 (+9/-6) | 2023/11 | |
| GPT-4o-2024-05-13 | 1209 (+6/-6) | OpenAI | 2023/10 |
| Gemini-1.5-Flash-Exp-0827 | 1208 (+11/-12) | 2023/11 | |
| Claude 3.5 Sonnet | 1191 (+6/-4) | Anthropic | 2024/4 |
| Gemini-1.5-Pro-001 | 1151 (+8/-6) | 2023/11 | |
| GPT-4-Turbo-2024-04-09 | 1151 (+7/-4) | OpenAI | 2023/12 |
| GPT-4o-mini-2024-07-18 | 1120 (+6/-5) | OpenAI | 2023/10 |
| Gemini-1.5-Flash-8b-Exp-0827 | 1110 (+9/-10) | 2023/11 | |
| Qwen2-VL-72B | 1085 (+26/-19) | Alibaba | Unknown |
| Claude 3 Opus | 1075 (+5/-6) | Anthropic | 2023/8 |
| Gemini-1.5-Flash-001 | 1072 (+7/-6) | 2023/11 | |
| InternVL2-26b | 1068 (+8/-7) | OpenGVLab | 2024/7 |
| Claude 3 Sonnet | 1048 (+6/-6) | Anthropic | 2023/8 |
| Yi-Vision | 1039 (+15/-15) | 01 AI | 2024/7 |
| qwen2-vl-7b-instruct | 1037 (+23/-21) | Alibaba | Unknown |
| Reka-Flash-Preview-20240611 | 1024 (+8/-6) | Reka AI | Unknown |
| Reka-Core-20240501 | 1015 (+5/-6) | Reka AI | Unknown |
| InternVL2-4b | 1010 (+9/-8) | OpenGVLab | 2024/7 |
| LLaVA-v1.6-34B | 1000 (+9/-7) | LLaVA | 2024/1 |
| Claude 3 Haiku | 1000 (+7/-6) | Anthropic | 2023/8 |
| LLaVA-OneVision-qwen2-72b-ov-sft | 992 (+16/-13) | LLaVA | 2024/8 |
| CogVLM2-llama3-chat-19b | 990 (+13/-12) | Zhipu AI | 2024/7 |
| MiniCPM-v 2_6 | 976 (+15/-13) | OpenBMB | 2024/7 |
| Phi-3.5-vision-instruct | 916 (+11/-10) | Microsoft | 2024/8 |
| Phi-3-Vision-128k-Instruct | 874 (+15/-12) | Microsoft | 2024/3 |
The updated leaderboard continues to evaluate a wide spectrum of models from renowned AI research organizations such as OpenAI, Anthropic, Google, Meta, and Reka AI. It provides a comprehensive overview of model performance, considering metrics like Arena Elo rating, MT-bench score, and MMLU score. This latest update underscores the ongoing competition and rapid innovation in AI, with new models consistently pushing the boundaries of performance benchmarks. As of September 2024, the LMSYS Chatbot Arena Leaderboard remains an essential resource for tracking the state-of-the-art in LLM capabilities, offering valuable insights into the evolving landscape of artificial intelligence.
How does the LMSYS Chatbot Arena Leaderboard work?
The LMSYS Chatbot Arena Leaderboard evaluates LLMs through a combination of user feedback and automated scoring systems. Participants can engage with chatbots and provide feedback, which is then used to calculate the Arena Elo rating. This process ensures that the leaderboard reflects both human preferences and objective performance metrics.
The leaderboard is an essential resource for developers and researchers, offering insights into the strengths and weaknesses of various models. It helps identify areas for improvement and guides the development of more advanced conversational AI systems.
What is the purpose of the LMSYS Chatbot Arena Leaderboard?
The purpose of the LMSYS Chatbot Arena Leaderboard is to provide a transparent and dynamic evaluation of LLMs in conversational settings. By incorporating user feedback and automated scoring, it offers a comprehensive view of model performance, helping to drive innovation and improvement in AI technology.
The leaderboard is designed to foster collaboration and knowledge sharing among AI researchers and developers, promoting the development of more effective and engaging conversational models.
Future Directions for the LMSYS Chatbot Arena Leaderboard
Future directions for the LMSYS Chatbot Arena Leaderboard include expanding the range of evaluation metrics, incorporating more diverse conversational scenarios, and enhancing user engagement. By continuously evolving, the leaderboard aims to remain at the forefront of AI evaluation, providing valuable insights into the capabilities of the latest LLMs.
Current Arena successor context
The following material is additive and documents the successor project's current snapshots and revised methodology.
The LMSYS Chatbot Arena Leaderboard was a ranking platform that assessed the performance of large language models (LLMs) in conversational tasks through blind human pairwise comparisons and statistical rating models.
The project began as an academic effort from LMSYS (Large Model Systems Organization) at UC Berkeley. It was later rebranded LMArena in 2024 and then simply Arena, and now operates as an independent company (arena.ai) rather than a university research project. The snapshots below use the successor project's July 2026 data.
Key features of the leaderboard included:
- Diverse Evaluation Metrics — The leaderboard used multiple metrics for evaluation, including Chatbot Arena Elo, which was based on user feedback, alongside other performance indicators such as MT-bench.
- Frequent Updates — Rankings were updated as new models were added to the arena, so any static snapshot ages quickly.
- Community Engagement — Users could participate in the evaluation process by voting on head-to-head chatbot responses, contributing to the dynamic nature of the rankings.
Arena Text Overall Snapshot (July 14, 2026)
This text ranking is the style-controlled Overall view in the official text data pinned to its July 14, 2026 revision. Style control reduces the influence of response-style features on preference ratings, while the confidence interval communicates statistical uncertainty.
| Rank | Model | Rating | 95% confidence interval | Votes |
|---|---|---|---|---|
| 1 | claude-fable-5 | 1507.5 | 1500.1–1515.0 | 7,959 |
| 2 | claude-opus-4-6-thinking | 1503.6 | 1499.8–1507.4 | 59,871 |
| 3 | claude-opus-4-7-thinking | 1502.9 | 1498.6–1507.2 | 47,141 |
| 4 | claude-opus-4-6 | 1497.7 | 1494.0–1501.4 | 63,636 |
| 5 | claude-opus-4-7 | 1494.1 | 1489.8–1498.4 | 48,248 |
Arena Vision Overall Snapshot (July 12, 2026)
The vision ranking is a separate style-controlled Overall snapshot from the official vision data pinned to its July 12, 2026 revision. It reflects human preferences on image-and-text conversations, so its ratings should not be compared numerically with the text table.
| Rank | Model | Rating | 95% confidence interval | Votes |
|---|---|---|---|---|
| 1 | claude-fable-5 | 1318.1 | 1307.4–1328.8 | 4,383 |
| 2 | claude-opus-4-7-thinking | 1304.0 | 1296.9–1311.2 | 18,567 |
| 3 | claude-opus-4-6-thinking | 1299.3 | 1292.2–1306.3 | 18,430 |
| 4 | claude-opus-4-7 | 1298.6 | 1291.5–1305.6 | 19,049 |
| 5 | claude-opus-4-6 | 1298.4 | 1291.6–1305.1 | 22,778 |
Both tables are dated snapshots of a dynamic preference leaderboard. Rankings can move as votes accumulate, models enter or leave the arena, and the statistical methodology changes. The Arena leaderboard publishes the continuously updated view.
Historical LMSYS Text Snapshot (repository archive: September 25, 2024)
The repository page archived on September 25, 2024 contained the following broader LMSYS Chatbot Arena text ranking. The old table mixed Arena ratings with mislabeled MT-Bench, MMLU, uncertainty, vote-count, and release-date fields, so only the model names and Arena ratings that can be coherently recovered are preserved here. The archive did not record a versioned leaderboard export or the precise ranking-method revision.
| Rank | Model | Arena rating |
|---|---|---|
| 1 | o1-preview | 1355 |
| 2 | ChatGPT-4o-latest (2024-09-03) | 1335 |
| 3 | o1-mini | 1324 |
| 4 | Gemini-1.5-Pro-Exp-0827 | 1299 |
| 5 | Grok-2-08-13 | 1294 |
| 6 | GPT-4o-2024-05-13 | 1285 |
| 7 | GPT-4o-mini-2024-07-18 | 1273 |
| 8 | Claude 3.5 Sonnet | 1269 |
| 9 | Gemini-1.5-Flash-Exp-0827 | 1269 |
| 10 | Grok-2-Mini-08-13 | 1267 |
| 11 | Gemini Advanced App (2024-05-14) | 1267 |
| 12 | Meta-Llama-3.1-405b-Instruct-fp8 | 1266 |
| 13 | Meta-Llama-3.1-405b-Instruct-bf16 | 1264 |
| 14 | GPT-4o-2024-08-06 | 1263 |
| 15 | Gemini-1.5-Pro-001 | 1259 |
| 16 | GPT-4-Turbo-2024-04-09 | 1257 |
| 17 | GPT-4-1106-preview | 1251 |
| 18 | Mistral-Large-2407 | 1250 |
| 19 | Athene-70b | 1250 |
| 20 | Meta-Llama-3.1-70b-Instruct | 1249 |
| 21 | Claude 3 Opus | 1248 |
| 22 | GPT-4-0125-preview | 1245 |
| 23 | Yi-Large-preview | 1240 |
| 24 | Gemini-1.5-Flash-001 | 1227 |
| 25 | Deepseek-v2-API-0628 | 1219 |
| 26 | Gemma-2-27b-it | 1218 |
| 27 | Yi-Large | 1212 |
| 28 | Gemini App (2024-01-24) | 1209 |
| 29 | Nemotron-4-340B-Instruct | 1209 |
| 30 | GLM-4-0520 | 1207 |
| 31 | Llama-3-70b-Instruct | 1206 |
| 32 | Claude 3 Sonnet | 1201 |
Historical LMSYS Vision Snapshot (repository archive: September 25, 2024)
The same repository archive included this separate vision-conversation ranking. Its displayed ratings and asymmetric uncertainty values are retained, while unrelated knowledge-cutoff metadata has been omitted because it was not part of the Arena evaluation. The archive did not specify whether its uncertainty values use the same confidence construction as the current dataset.
| Rank | Model | Arena rating | Reported uncertainty |
|---|---|---|---|
| 1 | Gemini-1.5-Pro-Exp-0827 | 1231 | +9/-6 |
| 2 | GPT-4o-2024-05-13 | 1209 | +6/-6 |
| 3 | Gemini-1.5-Flash-Exp-0827 | 1208 | +11/-12 |
| 4 | Claude 3.5 Sonnet | 1191 | +6/-4 |
| 5 | Gemini-1.5-Pro-001 | 1151 | +8/-6 |
| 6 | GPT-4-Turbo-2024-04-09 | 1151 | +7/-4 |
| 7 | GPT-4o-mini-2024-07-18 | 1120 | +6/-5 |
| 8 | Gemini-1.5-Flash-8b-Exp-0827 | 1110 | +9/-10 |
| 9 | Qwen2-VL-72B | 1085 | +26/-19 |
| 10 | Claude 3 Opus | 1075 | +5/-6 |
| 11 | Gemini-1.5-Flash-001 | 1072 | +7/-6 |
| 12 | InternVL2-26b | 1068 | +8/-7 |
| 13 | Claude 3 Sonnet | 1048 | +6/-6 |
| 14 | Yi-Vision | 1039 | +15/-15 |
| 15 | qwen2-vl-7b-instruct | 1037 | +23/-21 |
| 16 | Reka-Flash-Preview-20240611 | 1024 | +8/-6 |
| 17 | Reka-Core-20240501 | 1015 | +5/-6 |
| 18 | InternVL2-4b | 1010 | +9/-8 |
| 19 | LLaVA-v1.6-34B | 1000 | +9/-7 |
| 20 | Claude 3 Haiku | 1000 | +7/-6 |
| 21 | LLaVA-OneVision-qwen2-72b-ov-sft | 992 | +16/-13 |
| 22 | CogVLM2-llama3-chat-19b | 990 | +13/-12 |
| 23 | MiniCPM-v 2_6 | 976 | +15/-13 |
| 24 | Phi-3.5-vision-instruct | 916 | +11/-10 |
| 25 | Phi-3-Vision-128k-Instruct | 874 | +15/-12 |
These historical ratings predate the current style-controlled text and vision datasets. Changes in vote pools, model aliases, category filters, statistical estimation, style control, and uncertainty calculation mean the 2024 and 2026 numbers are not directly comparable. Manual maintenance should locate the exact September 2024 LMSYS exports, recover vote counts and intervals only from coherent source fields, document the historical methodology revision, and refresh the broader tables from pinned current datasets when full-row coverage is needed.
How did the LMSYS Chatbot Arena Leaderboard work?
The leaderboard evaluated LLMs through blind user voting. Participants compared responses from two anonymized models side by side and voted for the better one, and those preferences were aggregated into statistical ratings with uncertainty intervals.
This approach carried over to Arena, the leaderboard's successor: it remains a resource for developers and researchers to compare model strengths and weaknesses and track progress in conversational AI.
What was the purpose of the LMSYS Chatbot Arena Leaderboard?
The purpose of the LMSYS Chatbot Arena Leaderboard was to provide a transparent, crowd-sourced evaluation of LLMs in conversational settings. By aggregating user preferences into statistical ratings, it offered a comparative view of model performance that helped drive innovation and improvement across the industry.
The leaderboard also fostered collaboration and knowledge sharing among AI researchers and developers, encouraging the development of more effective and engaging conversational models.
From LMSYS to Arena
The Chatbot Arena project has since evolved beyond its original academic scope: it was rebranded LMArena in 2024, later shortened to Arena, and now operates as an independent company (arena.ai) rather than a UC Berkeley research project. The July 2026 tables above preserve exact published values from the successor project's official dataset.
More terms
Continue exploring the glossary.
Glossary term
What is Data Labeling in Machine Learning?
It's time to build
Collaborate with your team on reliable Generative AI features.
Want expert guidance? Book a 1:1 onboarding session from your dashboard.