Perplexity boosts Sonar search models and drops citation fees

Running search-augmented AI in production usually comes with an unwelcome surprise: unpredictable token bills driven by massive web context.

Perplexity just gave its Sonar API family a major tune-up. The company rolled out tiered search modes, claimed performance wins over OpenAI, and dropped one of the most frustrating hidden fees in AI search.

Three search gears and free citation tokens

The update applies to Sonar, Sonar Pro, and Sonar Reasoning Pro, leaving only the specialized Sonar Deep Research model on the old setup.

Developers can now toggle between three search intensities: High, Medium, and Low. Low keeps the existing baseline pricing and focuses on cost-efficient, straightforward lookups.

Medium balances speed and depth for everyday queries. High throws maximum compute and search depth at the problem, designed for complex, multi-hop research where factual precision cannot slip.

Even better, Perplexity scrapped charges for citation tokens entirely across the new modes. When Sonar pulls excerpts and links to back up its answers, those source tokens will no longer show up on your invoice.

Benchmarking the High modes, Perplexity claims base Sonar beats GPT-4o Mini on factual accuracy while closing in on full GPT-4o with search. Sonar Pro supposedly outscores even pricier flagship alternatives at a fraction of the cost.

Existing integrations get a 30-day grace period where current billing stays default while teams opt in to test. After April 18, 2025, Sonar Pro and Sonar Reasoning Pro responses will stop returning citation token counts or search result tallies in the usage payload.

Why it matters

If you build RAG pipelines or customer-facing search bots, this update solves two major operational headaches: cost volatility and verification friction.

Waiving citation fees is a brilliant move. Penalizing developers for citing sources always felt counterproductive, especially when hallucination reduction is the entire point of grounded search. Now teams can let models reference as many sources as needed without watching their API meters spin out of control.

It also puts direct pressure on OpenAI and Google. While search-enabled GPT-4o remains expensive and opaque, Perplexity is turning its core search infrastructure into a flexible, enterprise-grade utility that undercuts frontier pricing.

There is one trade-off worth noting. Dropping citation tokens and search result counts from the API usage field simplifies invoices, but it strips away observability for teams that rely on those exact metrics to evaluate retrieval performance. Still, for anyone looking to build reliable AI search on a budget, Sonar just became much harder to ignore.