arena.ai · benchmark source
Arena
Live human-preference and agent boards from Arena (LMArena). Scores are board-specific — Text Arena Score, WebDev Arena Score, and Agent net improvement are not interchangeable.
Agent systems
Agent net improvement
A net-improvement percentage reported for agent systems on the Arena Agent leaderboard.
Why it matters
It provides an agent-specific external signal without presenting a percentage improvement as an Elo rating.
Use it when
You are comparing agent-system results on the Arena Agent board and understand the metric is not a general model rating.
The limitation
It reflects the Arena Agent board and evaluation mix, not factual accuracy, cost, or enterprise safety.
Leaderboard snapshot
Agent net improvement results in a readable view
Use the tabs for the primary score and other published fields from this board. The source link remains authoritative for the live table.
Local leader
Claude Opus 5 (High)
+12.19%
Rows shown
48
Net improvement
Snapshot date
2026-08-14
1 day old · not a live API feed
Score profile
Net improvement by published row
Higher is better in this view
Claude Opus 5 (High)Claude Fable 5 (High)Claude Opus 5 (Max)
GPT 5.6 Sol (xHigh)
Kimi K3 (Max)Claude Opus 4.8 (Thinking)GPT 5.5 (xHigh)Claude Opus 4.7 (Thinking)Claude Opus 4.7GPT 5.5 (High)Claude Sonnet 5 (High)Claude Opus 4.6
GLM 5.2 (Max)GPT 5.5
Grok 4.5GPT 5.4 (High)GPT 5.6 Luna (xHigh)
Deepseek V4 Flash (High) (20260731)
Gemini 3.7 Flash (High)GPT 5.6 Terra (xHigh)Claude Sonnet 4.6Claude Opus 4.8
Muse Spark 1.1Kimi K2.7 CodeGLM 5.1
Qwen3.7 MaxDeepSeek V4 ProGemini 3.5 Flash (High)Gemini 3.1 Pro PreviewKimi K2.6
Hy3Qwen3.7 Plus
Mimo V2.5 ProDeepSeek V4 FlashGemini 3.6 Flash
Minimax M3Gemini 3.5 Flash (Medium)
Inkling
Mistral Medium 3.5Grok 4.3 (High)Gemini 3 FlashGrok Build 0.1Gemini 3.5 Flash-LiteMinimax M2.7
Solar Pro 4
Nemotron 3 UltraGrok 4.3Gemma 4 31B| Rank | Model or system | Net improvement | Other details | Note |
|---|---|---|---|---|
| #1 | Claude Opus 5 (High)Anthropic | +12.19% | Board: Arena Agent | — |
| #2 | Claude Fable 5 (High)Anthropic | +12.01% | Board: Arena Agent | — |
| #3 | Claude Opus 5 (Max)Anthropic | +11.95% | Board: Arena Agent | — |
| #4 | GPT 5.6 Sol (xHigh)OpenAI | +10.72% | Board: Arena Agent | — |
| #5 | Kimi K3 (Max)Moonshot AI | +10.43% | Board: Arena Agent | — |
| #6 | Claude Opus 4.8 (Thinking)Anthropic | +9.54% | Board: Arena Agent | — |
| #7 | GPT 5.5 (xHigh)OpenAI | +8.70% | Board: Arena Agent | — |
| #8 | Claude Opus 4.7 (Thinking)Anthropic | +8.17% | Board: Arena Agent | — |
| #9 | Claude Opus 4.7Anthropic | +7.63% | Board: Arena Agent | — |
| #9 | GPT 5.5 (High)OpenAI | +7.63% | Board: Arena Agent | — |
| #11 | Claude Sonnet 5 (High)Anthropic | +7.37% | Board: Arena Agent | — |
| #12 | Claude Opus 4.6Anthropic | +6.70% | Board: Arena Agent | — |
| #13 | GLM 5.2 (Max)Z.ai | +6.67% | Board: Arena Agent | — |
| #14 | GPT 5.5OpenAI | +6.23% | Board: Arena Agent | — |
| #15 | Grok 4.5SpaceXAI | +5.73% | Board: Arena Agent | — |
| #16 | GPT 5.4 (High)OpenAI | +4.99% | Board: Arena Agent | — |
| #17 | GPT 5.6 Luna (xHigh)OpenAI | +4.29% | Board: Arena Agent | — |
| #18 | Deepseek V4 Flash (High) (20260731)DeepSeek | +3.93% | Board: Arena Agent | — |
| #19 | Gemini 3.7 Flash (High)Google | +3.61% | Board: Arena Agent | — |
| #20 | GPT 5.6 Terra (xHigh)OpenAI | +3.43% | Board: Arena Agent | — |
| #21 | Claude Sonnet 4.6Anthropic | +2.99% | Board: Arena Agent | — |
| #22 | Claude Opus 4.8Anthropic | +2.50% | Board: Arena Agent | — |
| #23 | Muse Spark 1.1Meta | +1.09% | Board: Arena Agent | — |
| #24 | Kimi K2.7 CodeMoonshot AI | +1.03% | Board: Arena Agent | — |
| #25 | GLM 5.1Z.ai | +0.53% | Board: Arena Agent | — |
| #26 | Qwen3.7 MaxQwen | -0.01% | Board: Arena Agent | — |
| #27 | DeepSeek V4 ProDeepSeek | -0.07% | Board: Arena Agent | — |
| #28 | Gemini 3.5 Flash (High)Google | -0.43% | Board: Arena Agent | — |
| #29 | Gemini 3.1 Pro PreviewGoogle | -0.58% | Board: Arena Agent | — |
| #30 | Kimi K2.6Moonshot AI | -0.63% | Board: Arena Agent | — |
| #31 | Hy3Tencent | -1.30% | Board: Arena Agent | — |
| #32 | Qwen3.7 PlusQwen | -1.81% | Board: Arena Agent | — |
| #33 | Mimo V2.5 ProXiaomi | -2.21% | Board: Arena Agent | — |
| #34 | DeepSeek V4 FlashDeepSeek | -2.33% | Board: Arena Agent | — |
| #35 | Gemini 3.6 FlashGoogle | -2.36% | Board: Arena Agent | — |
| #36 | Minimax M3MiniMax | -2.52% | Board: Arena Agent | — |
| #37 | Gemini 3.5 Flash (Medium)Google | -3.64% | Board: Arena Agent | — |
| #38 | InklingInclusionAI | -6.71% | Board: Arena Agent | — |
| #39 | Mistral Medium 3.5Mistral | -6.93% | Board: Arena Agent | — |
| #40 | Grok 4.3 (High)SpaceXAI | -8.47% | Board: Arena Agent | — |
| #41 | Gemini 3 FlashGoogle | -8.59% | Board: Arena Agent | — |
| #42 | Grok Build 0.1SpaceXAI | -9.04% | Board: Arena Agent | — |
| #43 | Gemini 3.5 Flash-LiteGoogle | -10.21% | Board: Arena Agent | — |
| #44 | Minimax M2.7MiniMax | -11.13% | Board: Arena Agent | — |
| #45 | Solar Pro 4Upstage | -12.10% | Board: Arena Agent | — |
| #46 | Nemotron 3 UltraNVIDIA | -14.56% | Board: Arena Agent | — |
| #47 | Grok 4.3SpaceXAI | -14.61% | Board: Arena Agent | — |
| #48 | Gemma 4 31BGoogle | -18.23% | Board: Arena Agent | — |
Score type
Arena Agent net improvement percentage
Method
Causal tracing across live agent sessions
Board
Arena Agent
Best use
Comparing orchestrator contribution on this board
Full Arena Agent net-improvement board mirrored from arena.ai/leaderboard/agent. Gemini 3.7 Flash (High) is the live net-improvement column (3.61% as of 2026-08-14), not the success-rate column (~9.9%). The live board updates continuously; confidence intervals often overlap below the top cluster, so treat mid-board order as directional.
Full Arena Agent net-improvement board mirrored from arena.ai/leaderboard/agent. Gemini 3.7 Flash (High) is the live net-improvement column (3.61% as of 2026-08-14), not the success-rate column (~9.9%). The live board updates continuously; confidence intervals often overlap below the top cluster, so treat mid-board order as directional.
Open Arena Agent leaderboardEvidence in this catalog
Where Agent net improvement fits
Closest verified examples we currently carry. A missing score is not a zero.
Model examples
- Claude Opus 5+12.19% net improvement (high)
Arena Agent snapshot retrieved 2026-08-14; mid-board order can move as sessions accumulate.
- Claude Fable 5+12.01% net improvement (high)
Arena Agent snapshot retrieved 2026-08-14.
- GPT-5.6 Sol+10.72% net improvement (xhigh)
Arena Agent snapshot retrieved 2026-08-14.
Harness examples
No directly comparable harness score is verified here yet.