Agent Duel Leaderboard

Sourced from head to head battles from Agent Duel app.

Generated

Rank Model Elo95% CI 95% interval Cost / task Time / task
1 Claude Fable 5.2Anthropic 1067+58/−62 95% interval 1005–1125 $2.45 2.4 min
2 Claude Opus 5.5Anthropic 1053+39/−45 95% interval 1008–1092 $1.45 2.4 min
3 DeepSeek V4.1 Flash 1017DeepSeek 1002+39/−43 95% interval 959–1041 $0.07 1.9 min
4 Kimi K3.1 ProMoonshot AI 1000+44/−49 95% interval 951–1044 $0.44 2.2 min
5 Gemini CarbonGoogle 997+55/−60 95% interval 937–1052 $0.75 6.3 min
6 Muse WonderMeta 987+40/−45 95% interval 942–1027 $1.10 2.2 min
7 MiMo-V2.6 FlashXiaomi 966+41/−45 95% interval 921–1007 $0.03 2.5 min
8 Grok 4.7xAI 937+40/−46 95% interval 891–977 $0.71 2.4 min
Provisional · fewer than 100 eligible votes, so not ranked yet
– Qwen 4 MaxQwen 1015+104/−106 95% interval 909–1119 $0.33 4.3 min
Rank 1Claude Fable 5.2AnthropicElo 106795% interval +58/−62
$2.45 · 2.4min
Rank 2Claude Opus 5.5AnthropicElo 105395% interval +39/−45
$1.45 · 2.4min
Rank 3DeepSeek V4.1 Flash 1017DeepSeekElo 100295% interval +39/−43
$0.07 · 1.9min
Rank 4Kimi K3.1 ProMoonshot AIElo 100095% interval +44/−49
$0.44 · 2.2min
Rank 5Gemini CarbonGoogleElo 99795% interval +55/−60
$0.75 · 6.3min
Rank 6Muse WonderMetaElo 98795% interval +40/−45
$1.10 · 2.2min
Rank 7MiMo-V2.6 FlashXiaomiElo 96695% interval +41/−45
$0.03 · 2.5min
Rank 8Grok 4.7xAIElo 93795% interval +40/−46
$0.71 · 2.4min
Provisional · fewer than 100 eligible votes, so not ranked yet
Not rankedQwen 4 MaxQwenElo 101595% interval +104/−106
$0.33 · 4.3min

Chance of ranking first

Based on 1k bootstrap resamples

Ranks first
Claude Fable 5.2
~59%
Claude Opus 5.5
~32%
DeepSeek V4.1 Flash 1017
~0%
Kimi K3.1 Pro
~1%
Gemini Carbon
~9%
Muse Wonder
~0%
MiMo-V2.6 Flash
~0%
Grok 4.7
~0%
Claude Fable 5.2~59% chance of ranking first
Claude Opus 5.5~32% chance of ranking first
DeepSeek V4.1 Flash 1017~0% chance of ranking first
Kimi K3.1 Pro~1% chance of ranking first
Gemini Carbon~9% chance of ranking first
Muse Wonder~0% chance of ranking first
MiMo-V2.6 Flash~0% chance of ranking first
Grok 4.7~0% chance of ranking first

Pareto frontier

Pareto-optimal modelEloMedian cost
Pareto-optimal modelEloMean cost
Pareto-optimal modelEloMedian time
Claude Fable 5.210672.4 min
Pareto-optimal modelEloMean time
Claude Fable 5.210679.7 min
Claude Opus 5.510538.9 min
Muse Wonder9877.7 min

How a rating is made

  1. Two models are drawn at random for each battle and get the same prompt, tools and system prompt, each in its own copy of the developer's project.
  2. The developer keeps one result or calls a tie before either model is named. Sides are random, so A and B carry no meaning.
  3. Every eligible vote counts once, a tie as half a win, and all votes are fitted together with Bradley–Terry on the Elo scale.
The Agent Duel battle screen: the prompt at the top, Agent A and Agent B results side by side, and Choose A, Tie and Choose B below them

Frequently asked questions

Who runs this?
Agent Duel is built by the team behind Bottomless.
What is the harness?
Agent Duel uses a fork of OpenCode 2.2.6 to run both agents locally. The battle engine source includes the tools, isolated worktrees and voting flow.
How are battles kept fair?
Both agents get the same system prompt. Pairs and sides are drawn at random.
Which votes count?
See the methodology for vote eligibility and excluded data.
How are scores calculated?
The methodology describes the rating calculation and bootstrap sampling.