Perplexity Sonar ties for top spot in new Search Arena

Search benchmarks love to measure simple trivia, but throw real humans with messy queries at AI search tools, and things get wild.

LM Arena just dropped its new Search Arena leaderboard, evaluating search-augmented LLMs based on over 10,000 blind human preference votes across 11 models.

Perplexity's Sonar models didn't just show up—they dominated.

Perplexity ties Google, clears OpenAI

Evaluated between March 18 and April 13, 2025, Search Arena ditched trivia questions in favor of real-world coding, research, writing, and complex current events.

Perplexity's Sonar-Reasoning-Pro-High snagged an Arena score of 1136, statistically tying Google's heavyweight Gemini-2.5-Pro-Grounding (1142) for the #1 spot.

In direct head-to-head battles, Sonar actually edged out Gemini-2.5-Pro 53% of the time.

Even wilder: Perplexity models locked down ranks one through four on the overall leaderboard, completely outclassing Gemini-2.0-Flash-Grounding and wiping the floor with OpenAI's entire web search lineup, including GPT-4o Search.

The secret sauce: Going deeper on sources

Why did human judges consistently pick Sonar? The data reveals three big correlations with user preference: longer answers, community web sources, and way more citations.

Reasoning was the major differentiator, with reasoning-backed models sweeping the top three leaderboard spots.

Search depth turned out to be the real equalizer. Sonar models cited an average of 2x to 3x more sources than comparable Gemini models.

Interestingly, researchers ran a control test: when they artificially limited citation counts, the performance gaps between models basically vanished. In short, Sonar is winning because it actively digs through more pages before serving up an answer.

Why it matters

If you rely on AI for research, heavy documentation, or market intelligence, this is clear validation that Perplexity's reasoning search models are the current gold standard. Setting Sonar as your default or calling it via API gives you significantly better grounded answers than generic frontier chat models.

For the AI industry, it's a huge strategic win for specialized product engineering. Google and OpenAI have virtually unlimited compute and massive base models, but Perplexity beat them in their own backyard simply by building a better retrieval-and-reasoning loop.

We think Perplexity's focus on deep citation and reasoning is the exact right bet. Users don't just want confident text; they want receipts. That said, pulling 3x more sources and running multi-step reasoning adds latency and inference cost. The real challenge now is keeping that search depth without making users wait an eternity for a quick answer.