Skip to main content
AI Choice Engine
← Benchmark sources

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. The Arena product home at arena.ai/ and the long-running lmarena.ai leaderboards are the same operator (verified 2026-08-29 via arena.ai root and the LMArena research pages; see https://lmarena.ai/ for the historic domain).

Open arena.ai3 boards on this publisher

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.

Capture updated October 1 from the September 30 board. Negative net-improvement values mean the agent hurt outcomes versus baseline; mid-board order can shift as sessions accumulate and confidence intervals overlap.
Arena Agent net-improvement capture (48 rows, board dated 2026-09-30). Top: Claude Fable 5.1 (Max) (14.55%), Claude Opus 5.5 (High) (13.78%), GPT 6 Astra (Max) (12.18%).

Leader among verified rows

Claude Fable 5.1 (Max)

+14.55%

Rows shown

48

Net improvement

Source / snapshot date

2026-10-01

25 days old · not a live API feed

Score profile

Net improvement by published row

Higher is better in this view

Claude Fable 5.1 (Max)
+14.55%
Claude Opus 5.5 (High)
+13.78%
GPT 6 Astra (Max)
+12.18%
GPT 6 Sol (Max)
+10.65%
Claude Opus 5 (High)
+8.76%
Claude Opus 5 (Max)
+8.55%
Claude Fable 5 (High)
+8.37%
Gemini 4 Argon (High)
+7.92%
GPT 5.6 Sol (xHigh)
+7.05%
Claude Opus 4.8 (High)
+6.92%
Claude Sonnet 5 (High)
+4.81%
Hy4 preview
+4.24%
Muse Spark 1.3 (Max)
+4.19%
GPT 5.5 (xHigh)
+4.12%
Deepseek V4.1 Flash (Max)
+4.01%
Kimi K3 (Max)
+3.98%
GLM 5.2 (Max)
+3.64%
Grok 4.7 (xHigh)
+3.64%
MiMo V2.6 Pro
+3.17%
Gemini 3.8 Flash (High)
+2.96%
Qwen3.8 Max
+2.80%
GLM 5.3 (Max)
+2.57%
GPT 6 Luna (Max)
+1.68%
DeepSeek V4 Pro (High) (0813)
+1.29%
Grok 4.6 (xHigh)
+1.16%
GPT 5.5
+1.13%
Grok 4.5
+1.11%
GPT 5.4 (High)
+0.91%
GPT 5.6 Terra (xHigh)
+0.83%
GLM 5.3 Flash
-0.28%
MiMo V2.6 Flash
-0.57%
Qwen3.8 Flash Next
-0.79%
GPT 5.6 Luna (xHigh)
-1.09%
Qwen 3.8 27B
-1.61%
Gemini 3.7 Flash (High)
-1.64%
Muse Spark 1.2 (xHigh)
-2.96%
Muse Spark 1.1
-4.60%
Qwen3.7 Max
-5.20%
Hy3
-5.70%
Minimax M3
-6.71%
Mimo V2.5 Pro
-7.23%
Qwen3.7 Plus
-7.28%
Gemini 3.1 Pro Preview
-7.35%
Gemini 3.6 Flash (High)
-7.67%
Inkling Small
-9.80%
Inkling
-10.90%
Mistral Medium 3.5
-11.74%
Solar Pro 4
-15.77%

Full Arena Agent net-improvement table captured 2026-10-01 from arena.ai/leaderboard/agent (48 models, 2,022,521 sessions; board dated 2026-09-30). Score signs come from the accessible net-improvement labels; rows also carry the displayed confidence interval, session count, P50 task cost and P50 output tokens. Confidence intervals overlap across much of the board, so treat close ranks as directional.

Open Arena Agent leaderboard

Evidence in this catalog

Where Agent net improvement fits

Closest verified examples we currently carry. A missing score is not a zero.

Model examples

  • Published source observation captured 2026-10-01. Source-specific result; no independent rerun or universal capability verdict.

  • Published source observation captured 2026-10-01. Source-specific result; no independent rerun or universal capability verdict.

  • Published source observation captured 2026-10-01. Source-specific result; no independent rerun or universal capability verdict.

Harness examples

No directly comparable harness score is verified here yet.

Open this board on Arena Agent leaderboard