BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

Arena Agent

Task-outcome score (0–100) from real agentic sessions judged by their users.

As of 24 Oct 2026, Claude Opus 5.5 leads Arena Agent on BenchLeader with 20.9, ahead of GPT-6 Astra at 14.7, across 50 model configurations with a published result.

Published by
Arena (LMArena)
Category
Agents & tools
Index weight
0.5
Models
50
Data as of
24 Oct 2026
Newest result
20 Oct 2026

CC BY 4.0 — lmarena-ai/leaderboard-dataset on Hugging Face.

What the test looks like

The Agent arena logs real agentic sessions, where a model works through a multi-step task with tools, and asks the person who ran it to judge the outcome.

How it is scored

A task-outcome score from 0 to 100 based on user judgments of completed sessions, rather than an Elo rating.

What to keep in mind

New and lightly sampled; sessions vary in length and difficulty, and the user is both the task setter and the judge.

Top 15
  1. 1Claude Opus 5.5 (max)20.9
  2. 2GPT-6 Astra (max)14.7
  3. 3Claude Fable 5.1 (max)14.2
  4. 4GPT-6.1 Sol (max)13.5
  5. 5Claude Sonnet 5.5 (max)12.7
  6. 6Gemini 4 Argon (high)12
  7. 7GPT-6 Sol (max)9.4
  8. 8Claude Fable 5 (high)8.3
  9. 9Claude Opus 5 (max)8.1
  10. 10GPT-5.6 Sol (xhigh)6.6
  11. 11Claude Sonnet 5 (high)4.2
  12. 12Muse Wonder (max)3.7
  13. 13MiMo-V2.6-Pro3.7
  14. 14Gemini 3.8 Flash (high)3.6
  15. 15Muse Spark 1.3 (max)3.4
50 of 50
#
1Claude Opus 5.5max20.972.62026-10-20
2GPT-6 Astramax14.770.42026-10-20
3Claude Fable 5.1max14.270.52026-10-20
4GPT-6.1 Solmax13.569.92026-10-20
5Claude Sonnet 5.5max12.769.32026-10-20
6Gemini 4 Argonhigh1270.82026-10-20
7GPT-6 Solmax9.466.42026-10-20
8Claude Fable 5high8.364.62026-10-20
9Claude Opus 5max8.168.82026-10-20
10GPT-5.6 Solxhigh6.666.62026-10-20
11Claude Sonnet 5high4.260.22026-10-20
12Muse Wondermax3.767.52026-10-20
13MiMo-V2.6-Pro3.764.02026-10-20
14Gemini 3.8 Flashhigh3.664.42026-10-20
15Muse Spark 1.3max3.467.12026-10-20
16Grok 4.7xhigh3.261.62026-10-20
17GPT-5.5xhigh3.166.12026-10-20
18GPT-6 Lunamax360.02026-10-20
19DeepSeek V4.1 Flashmax2.661.22026-10-20
20MiMo-V2.6-Flash2.559.72026-10-20
21GLM 5.2max2.162.32026-10-20
22Kimi K3max1.864.72026-10-20
23Qwen3.8 Maxmax1.663.32026-10-20
24Claude Haiku 5.5high1.559.62026-10-20
25GLM 5.3max1.463.52026-10-20
26GPT-5.4high1.457.92026-10-20
27Grok 4.6xhigh0.762.52026-10-20
28GPT-5.6 Terraxhigh0.562.82026-10-20
29DeepSeek V4 Prohigh0.157.52026-10-20
30GPT-5.5-0.560.32026-10-20
31Gemini 3.7 Flashhigh-163.52026-10-20
32Step 5 Preview-1.560.42026-10-20
33GLM 5.3 Flash-1.662.62026-10-20
34GPT-5.6 Lunaxhigh-2.359.42026-10-20
35Qwen3.8-Flash-Next-3.160.22026-10-20
36Gemini 3.1 Pro-3.760.62026-10-20
37Qwen3.8 27B-3.854.62026-10-20
38NewBeam 1.0 501B A23B-4.454.42026-10-20
39Mistral Large 4-4.656.62026-10-20
40GPT-6 Lunalow-4.651.22026-10-20
41Qwen3.7max-6.260.62026-10-20
42Muse Spark 1.1-6.662.52026-10-20
43Qwen3.7 Plus-6.758.52026-10-20
44Muse Spark 1.2xhigh-6.863.22026-10-20
45MiMo-V2.5-Pro-8.254.92026-10-20
46MiniMax M3-9.656.42026-10-20
47Mistral Medium 3.5-11.849.42026-10-20
48Nemotron 3 Ultra-12.344.12026-10-20
49Inkling Small-15.854.82026-10-20
50Solar Pro 4-19.653.32026-10-20

More agents & tools benchmarks · All LLM benchmarks

Cite as: BenchLeader, “Arena Agent leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/lmarena_agent.

Arena Agent: questions

What does Arena Agent measure?
The Agent arena logs real agentic sessions, where a model works through a multi-step task with tools, and asks the person who ran it to judge the outcome. Scores are reported in score; higher is better.
Which AI model leads Arena Agent?
Claude Opus 5.5 leads Arena Agent with 20.9 as of 24 Oct 2026, ahead of GPT-6 Astra at 14.7.
How many models have Arena Agent results?
50 model configurations have a Arena Agent result on BenchLeader, all taken from Arena (LMArena).
Who runs Arena Agent and how often is it updated?
Arena Agent is published by Arena (LMArena). BenchLeader re-reads the published results once a day and records the date each result was published.
Does Arena Agent count toward the BenchLeader Index?
Yes. Arena Agent contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.