LMArena Leaderboard, September 2026
What the LMArena actually is, how to read an Arena Elo score, and the current top 10 for September 2026. The original human-preference benchmark that started as LMSys Chatbot Arena and now anchors most enterprise model selection conversations.
What is the LMArena, in one paragraph?
The LMArena is a public, blind side-by-side voting site for AI chat models. A user submits a prompt, two anonymous models reply, the user picks a winner, and the project aggregates millions of such votes into Elo ratings. It started in 2023 as the LMSys Chatbot Arena out of UC Berkeley and rebranded to LMArena.ai in 2024-25 as it spun out into an independent project. The current September 2026 top 10 is below: three models now sit above the historical 1500 Elo barrier on text, with the open-weights tier within striking distance of the closed-source frontier.
How to read an Arena Elo score
Reference table for Arena Elo bands (September 2026) 1510+ Frontier #1 Claude Opus 4.8 (AAII 61.4, coding & overall #1) 1500 Frontier Gemini 3.1 Pro, Claude Opus 4.7, GPT-5.5 Pro 1450 Frontier-adj. DeepSeek V4 Pro, Qwen 3.7 Max 1400 Strong tier GPT-4.1, Claude Sonnet 4, Gemini 2.5 Pro 1300 Capable tier Llama 4 Maverick, Mistral Large 3 1200 Solid daily Gemma 4, Phi-4, Mistral Small 3 1100 Light tasks DeepSeek V4 Flash, GPT-4o Mini <1100 Legacy tier Older 2023-24 model generations A 100-Elo gap means the higher-rated model wins ~64% of head-to-heads. A 200-Elo gap means it wins ~76%. Rating shifts under 25 points are noise.
Live Leaderboard
| # | Model | Quality | Arena ELO | Speed | Price | Context | Value | Released |
|---|---|---|---|---|---|---|---|---|
| 1 | Anthropic · Frontier agentic coding & knowledge work | 100 | 1525 | 58 t/s | $10 / $50 | 1M | 3.3 | Jun 2026 |
| 2 | Anthropic · Ceiling capability (limited access) | 100 | 1531 | 56 t/s | $10 / $50 | 1M | 3.3 | Jun 2026 |
| 3 | Anthropic · Frontier agentic coding & reasoning | 99 | 1522 | 74 t/s | $5 / $25 | 1M | 6.6 | Jul 2026 |
| 4 | Anthropic · Coding, agents & computer use | 98 | 1512 | 72 t/s | $5 / $25 | 1M | 6.5 | May 2026 |
| 5 | OpenAI · General frontier reasoning & tools | 98 | 1514 | 96 t/s | $5 / $30 | 1M | 5.6 | Jul 2026 |
| 6 | OpenAI · Frontier general purpose | 97 | 1506 | 70 t/s | $5 / $30 | 1M | 5.5 | Apr 2026 |
| 7 | Kimi K3OSS Moonshot AI · Open-weight coding frontier | 97 | 1500 | 55 t/s | $3 / $15 | 1M | 10.8 | Jul 2026 |
| 8 | OpenAI · Reasoning at any cost | 96 | 1510 | 68 t/s | $30 / $180 | 1M | 0.9 | Apr 2026 |
| 9 | Anthropic · Coding & agentic workflows | 96 | 1505 | 68 t/s | $5 / $25 | 1M | 6.4 | Apr 2026 |
| 10 | Google · Science & long-context | 96 | 1505 | 131 t/s | $2 / $12 | 1M | 13.7 | Apr 2026 |
| 11 | Qwen3.8 MaxOSS Alibaba Cloud · Multimodal APAC frontier (hosted); text-only open weights | 96 | 1491 | 86 t/s | $2 / $6 | 1M | 24.0 | Aug 2026 |
| 12 | OpenAI · Hard reasoning | 94 | 1370 | 68 t/s | $2 / $8 | 200K | 18.8 | Apr 2025 |
| 13 | Alibaba Cloud · Long autonomous agentic runs | 94 | 1488 | 90 t/s | $2.5 / $7.5 | 1M | 18.8 | May 2026 |
| 14 | xAI · Cheap output tokens & real-time info | 94 | 1499 | 88 t/s | $2 / $6 | 500K | 23.5 | Jul 2026 |
| 15 | xAI · Agentic tasks & real-time info | 93 | 1496 | 83 t/s | $1.25 / $2.5 | 1M | 49.6 | May 2026 |
| 16 | OpenAI · Default production tier | 93 | 1489 | 118 t/s | $2 / $12 | 1M | 13.3 | Jul 2026 |
| 17 | Anthropic · Balanced agents & coding value | 93 | 1479 | 98 t/s | $3 / $15 | 1M | 10.3 | Jun 2026 |
| 18 | Google · Multimodal + value | 92 | 1345 | 87 t/s | $1.25 / $10 | 1M | 16.4 | Mar 2025 |
| 19 | Moonshot AI · Frontier quality at low cost | 92 | 1466 | 48 t/s | $0.73 / $3.49 | 256K | 43.6 | Apr 2026 |
| 20 | Anthropic · Complex analysis | 91 | 1360 | 52 t/s | $15 / $75 | 200K | 2.0 | May 2025 |
| 21 | DeepSeek R1OSS DeepSeek · Cheap reasoning | 91 | 1350 | 35 t/s | $0.55 / $2.19 | 128K | 66.4 | Jan 2025 |
| 22 | GLM-5.2OSS Z.ai (Zhipu AI) · Fastest self-hostable frontier model | 91 | 1483 | 168 t/s | $1.4 / $4.4 | 1M | 31.4 | Jun 2026 |
| 23 | Anthropic · Coding & balance | 90 | 1467 | 73 t/s | $3 / $15 | 1M | 10.0 | Feb 2026 |
| 24 | Google · Fast multimodal at mid-tier price | 90 | 1481 | 148 t/s | $0.75 / $3.75 | 1M | 40.0 | Jul 2026 |
| 25 | OpenAI · Long context | 89 | 1310 | 120 t/s | $2 / $8 | 1M | 17.8 | Apr 2025 |
| 26 | DeepSeek · Cheap frontier-adjacent tokens (GA build) | 89 | 1461 | 62 t/s | $0.66 / $1.98 | 1M | 67.4 | Aug 2026 |
| 27 | MiniMax M3OSS MiniMax · Open-weight agentic coding | 89 | 1455 | 80 t/s | $0.6 / $2.4 | 1M | 59.3 | Jun 2026 |
| 28 | OpenAI · Reasoning & math | 88 | 1305 | 155 t/s | $1.1 / $4.4 | 200K | 32.0 | Jan 2025 |
| 29 | Anthropic · Coding & balance | 88 | 1320 | 95 t/s | $3 / $15 | 200K | 9.8 | May 2025 |
| 30 | GLM-5.1OSS Z.ai (Zhipu AI) · Open-weight agentic & tool use | 88 | 1467 | 48 t/s | $0.98 / $3.08 | 200K | 43.3 | Apr 2026 |
| 31 | xAI · Real-time info | 87 | 1330 | 82 t/s | $3 / $15 | 131K | 9.7 | Feb 2025 |
| 32 | DeepSeek V3OSS DeepSeek · Best open-source value | 86 | 1310 | 62 t/s | $0.27 / $1.1 | 128K | 125.5 | Mar 2025 |
| 33 | Alibaba Cloud · Multilingual & APAC | 86 | 1448 | 124 t/s | $1.4 / $5.6 | 256K | 24.6 | Apr 2026 |
| 34 | OpenAI · General purpose | 85 | 1285 | 109 t/s | $2.5 / $10 | 128K | 13.6 | May 2024 |
| 35 | DeepSeek · Cheapest tokens, period | 84 | 1441 | 105 t/s | $0.14 / $0.28 | 1M | 400.0 | Apr 2026 |
| 36 | OpenAI · High-volume cheap frontier-family | 83 | 1436 | 186 t/s | $0.2 / $1.2 | 1M | 118.6 | Jul 2026 |
| 37 | Meta · Open-source value | 80 | 1260 | 135 t/s | $0.2 / $0.6 | 1M | 200.0 | Apr 2025 |
| 38 | Qwen 2.5 72BOSS Alibaba Cloud · Open-source flagship | 80 | 1255 | 85 t/s | $0.3 / $0.9 | 131K | 133.3 | Sep 2024 |
| 39 | Qwen3.8 27BOSS Alibaba Cloud · Locally-runnable open-weight (~17GB VRAM at 4-bit) | 80 | — | — | Self-host | 262K | , | Aug 2026 |
| 40 | Mistral AI · Multilingual | 79 | 1250 | 78 t/s | $2 / $6 | 128K | 19.8 | Nov 2024 |
| 41 | Google · High-volume cheap multimodal | 79 | 1418 | 192 t/s | $0.3 / $2.5 | 1M | 56.4 | Jul 2026 |
| 42 | NVIDIA · Efficient single-GPU deployment, 1M context | 79 | — | — | Self-host | 1M | , | Aug 2026 |
| 43 | xAI · Budget reasoning | 78 | 1275 | 165 t/s | $0.3 / $0.5 | 131K | 195.0 | Feb 2025 |
| 44 | Perplexity · Search + citations | 78 | — | 65 t/s | $3 / $15 | 200K | 8.7 | Feb 2025 |
| 45 | Mistral AI · Code generation | 76 | — | 195 t/s | $0.3 / $0.9 | 256K | 126.7 | Jan 2025 |
| 46 | NVIDIA · Open multimodal | 76 | 1361 | 158 t/s | Self-host | 256K | , | Apr 2026 |
| 47 | Hunyuan Hy3OSS Tencent · Long-context retrieval & agentic search, no license restrictions | 76 | — | — | Self-host | 256K | , | Jul 2026 |
| 48 | Anthropic · Speed & cost | 75 | 1230 | 172 t/s | $0.8 / $4 | 200K | 31.3 | Oct 2024 |
| 49 | Gemma 4 27BOSS Google · Self-hosted general purpose | 75 | 1351 | 142 t/s | Self-host | 128K | , | Apr 2026 |
| 50 | Google · Fastest + cheapest | 74 | 1240 | 244 t/s | $0.1 / $0.4 | 1M | 296.0 | Feb 2025 |
| 51 | Alibaba Cloud · Open-source coding | 74 | — | 125 t/s | $0.15 / $0.45 | 131K | 246.7 | Nov 2024 |
| 52 | Ant Group · Efficient MoE alternative to trillion-param flagships | 74 | — | — | Self-host | 256K | , | Aug 2026 |
| 53 | OpenAI · High throughput | 72 | 1216 | 183 t/s | $0.15 / $0.6 | 128K | 192.0 | Jul 2024 |
| 54 | Meta · Longest context | 71 | 1195 | 198 t/s | $0.15 / $0.4 | 10M | 258.2 | Apr 2025 |
| 55 | Amazon · AWS ecosystem | 70 | — | 110 t/s | $0.8 / $3.2 | 300K | 35.0 | Dec 2024 |
| 56 | Cohere · Enterprise RAG | 68 | 1170 | 72 t/s | $2.5 / $10 | 128K | 10.9 | Aug 2024 |
What to do this quarter
- Treat Arena Elo as a triage filter, not a decision. Use it to drop the bottom half of your candidate list, then run a real eval on the remainder.
- Pick the right Arena board. Coding teams should read the coding Arena (Claude Opus 4.8 now leads at ~1582 Elo, ahead of Opus 4.7 at 1567). Long-context teams should read the hard-prompts Arena. The aggregate text leaderboard is the wrong signal for many enterprise workloads.
- Discount short-conversation polish. The Arena rewards style. Models tuned for chat win at the margin against models tuned for accuracy. Build internal evals that reward what your business actually pays for.
- Watch the gap, not the ranking. Sub-25 Elo shifts are within statistical noise. Anything under 50 Elo between two candidates is a coin flip on most workloads.
- Plan for the multi-way race. Claude Opus 4.8 holds a narrow lead, but Gemini 3.1 Pro, Claude Opus 4.7, GPT-5.5 Pro, Qwen 3.7 Max, and DeepSeek V4 Pro are approximately interchangeable on quality at the top. Optimise your stack for switching cost, not for capability.
- Capture vote-rate momentum. The fastest-rising models week-over-week are usually the next month's leaders. Subscribe to weekly Arena reports.
- Pair Arena Elo with cost. A 50-Elo lead at 10x the price is rarely a good trade. See our model leaderboard for combined quality-cost rankings.
Related reading
- AI Model Leaderboard: full quality, speed, and pricing comparison
- LLM Leaderboard: same data, LLM-focused entry point
- LMSys Arena Leaderboard May 2026 : full deep-dive
- LMArena Elo Explained for Enterprise Buyers
For teams running multiple top-of-Arena models in production, Swfte Connect provides a single OpenAI-compatible endpoint that routes across providers and normalises Arena-tier quality without re-architecting your stack.