Perplexity’s new Sonar model hits 1,200 tokens per second

The biggest friction with AI search engines has never been finding links. It is sitting there watching words trickle onto the screen like a slow faucet when all you want is a fast answer.

Perplexity just rolled out the latest version of Sonar, its in-house search-specialized model. Built on top of Meta's open-source Llama 3.3 70B, Sonar has been fine-tuned strictly for two things: factuality and readability.

To eliminate the waiting game entirely, Perplexity hooked it up to Cerebras inference infrastructure, clocking an eye-popping 1,200 tokens per second.

Instant answers at 1,200 tokens per second

To put that speed into perspective, Sonar delivers answers roughly 10 times faster than lightweight speedsters like Gemini 2.0 Flash. The moment you hit enter, your research summary is basically already there.

Speed means nothing if the output is garbage, but Perplexity claims Sonar punches well above its weight class. In extensive online A/B testing on user satisfaction, Sonar substantially beat rival lightweight models like GPT-4o mini and Claude 3.5 Haiku.

More impressively, it edged out frontier models like Claude 3.5 Sonnet and ran neck-and-neck with GPT-4o, while operating at a fraction of the cost.

Sonar also notched higher scores on standard benchmarks like IFEval for instruction following and MMLU for general knowledge compared to other models in its class.

Perplexity focused its post-training on resolving conflicting search results and structuring concise summaries with clean markdown. Pro subscribers can toggle Sonar on as their default model right now, while developers can access it via the Sonar API, though Cerebras-powered hardware for API traffic is still rolling out.

Why it matters

For everyday researchers and knowledge workers, this is an instant upgrade. Waiting on heavy frontier models for routine queries has always felt like overkill; Sonar gives you top-tier search synthesis with practically zero latency.

Strategically, Perplexity's playbook is brilliant. Rather than burning billions trying to train a frontier foundation model from scratch, it took an open-weight base in Llama 3.3 70B, tailored it ruthlessly for its exact workflow, and paired it with specialized Cerebras inference chips. It proves that vertical optimization can beat raw brute-force scale where it counts most: latency and unit economics.

We think this sets a serious precedent for application-layer AI companies. The risk, of course, is ceiling height. Sonar excels at factual grounding and crisp layout, but it won't replace o3-mini or Sonnet when you need deep multi-step reasoning or complex coding. But for conversational search, blistering speed paired with accurate citations is the real endgame.