Arena Search
Citation frequency 2/3
- What does it measure?
- Tests, through user votes, how good the answers are when a search-enabled AI replies to a question with citations.
- Who evaluates it, and how?
- A user asks a question; two anonymous search-enabled AIs answer with citations and the user votes for the better answer. The votes are fitted with a Bradley–Terry model, then converted to the Elo scale.
- Reading the score
- Scores are Elo ratings from head-to-head human votes. Higher is better. Citation formatting is standardized so the models cannot be recognized by style.
- Related benchmarks
- Arena Document|Arena Image-to-Video|Arena Text Factuality|Arena Video Edit
Top score1256
Models tested29
Scoring basisHigher is better
Last updated2026.11.19
Model release date
Which AI gives answers people prefer most, without false claims?Arena Text Factuality
Gemini 4 Carbon🥇
Gemini 4 Carbon🥇Arena Text Factuality
Gemini 4 Carbon🥇Does it know facts accurately without searching?SimpleQA Verified
Claude Fable 5.2🥇
Claude Fable 5.2🥇SimpleQA Verified
Claude Fable 5.2🥇Does it avoid making things up when summarizing?Vectara Hallucination Rate
GPT-5.4 Nano🥇
GPT-5.4 Nano🥇Vectara Hallucination Rate
GPT-5.4 Nano🥇
Which AI gives the search answers people prefer most?
Arena Search
- 🥇OpenAIGPT-5.6 Sol1256
- 🥈AnthropicClaude Opus 4.61252
- 🥉OpenAIGPT-5.51246
- #4AnthropicClaude Opus 5Reasoning effort: High1238
- #5AnthropicClaude Fable 5.1Reasoning effort: High1237
- #6OpenAIGPT-6 Astra1231
- #7AnthropicClaude Opus 4.71230
- #8OpenAIGPT-6.1 Sol1230
- #9AnthropicClaude Fable 51228
- #10AnthropicClaude Sonnet 4.61220
- #11SpaceXSIGrok 4.51216
- #12AnthropicClaude Opus 5.5Reasoning effort: Medium1214
- #13GoogleGemini 3.1 Pro1211
- #14AnthropicClaude Opus 4.81202
- #15OpenAIGPT-5.41198
- #16GoogleGemini 3 Flash1195
- #17OpenAIGPT-6.2 Astra1191
- #18SpaceXSIGrok 4.201191
- #19AnthropicClaude Sonnet 51187
- #20AnthropicClaude Opus 4.51179
- #21OpenAIGPT-6.2 Sol1179
- #22OpenAIGPT-6 Sol1174
- #23SpaceXSIGrok 4.31166
- #24AnthropicClaude Sonnet 4.51159
- #25OpenAIGPT-6 Luna1157
- #26AnthropicClaude Opus 4.11148
- #27GoogleGemini 2.5 Pro1143
- #28OpenAIGPT-51133
- #29AnthropicClaude Opus 41127
Top 3 within the margin of error · effectively tied · Half of models ≤ 1198
The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.