Parallel Web Systems' Turbo Cuts AI Search Costs 14x at 200ms
Parallel Web Systems launches Turbo mode for its Search API: 200ms median latency, $1 per 1,000 requests, and benchmark-leading accuracy against Exa, Tavily, and Brave.
- Parallel Search Turbo launches at $1/1,000 requests and 200ms median latency -- up to 14x cheaper than frontier model search. (Blog)
- Benchmark-leading accuracy: 51% on BrowseComp vs. Exa's 33.7%, Tavily's 19.3%, and Brave's 38.3%, all at faster latency.
- New architecture: Turbo is the first product on a re-engineered search stack spanning hardware, model training, and index design.
- Key use cases unlocked: voice AI, RL training with live web grounding, consumer chat with search-by-default, and small/local model augmentation.
- Company context: Parallel raised a $100M Series B at a $2B valuation led by Sequoia; total capital raised is $230M. (Series B announcement)
- Real-world traction: Nooks cut web search costs 70.5% switching to Parallel; Harvey, Notion, and Opendoor are also production customers.
Parallel Web Systems just launched Turbo, a new speed tier for its Search API that is gunning directly at the cost and latency assumptions that have defined AI web search. At $1 per 1,000 requests and a 200ms median latency, it is not a marginal improvement over the competition -- it is a different price point entirely.
The numbers that matter
The headline claims are specific and backed by benchmarks Parallel ran across four public datasets. On BrowseComp -- OpenAI's benchmark of 1,266 hard, multi-hop web questions -- Parallel Turbo achieved 51% accuracy at a 216ms median latency, compared to Exa Instant at 33.7% / 361ms, Tavily Ultra Fast at 19.3% / 357ms, Brave Search at 38.3% / 430ms, and SerpAPI at 23.3% / 999ms.
The cost story is equally sharp. With a median latency of 200ms and a price of $1 per 1,000 requests, Turbo is up to 14x cheaper than the default search in frontier models, while maintaining similar or better accuracy. For context, fast search APIs from competitors run $5--7 per 1,000 requests, frontier model search costs $10--14 per 1,000, and SERP APIs land around $1 per 1,000 but return raw results that require additional processing.
That last point matters. SERP APIs are cheap but dumb -- they hand you raw HTML and links. Turbo, like the rest of Parallel's Search API, returns dense, LLM-ready excerpts. By default, Turbo returns dense, relevant excerpts directly, so the model spends fewer input tokens on a better answer.
Why agents break the old search economics
The launch is timed to a structural shift in how software uses the web. A typical person runs a handful of searches a day, but an agent runs thousands: inside loops, across tools, and often multiple times per question. That volume makes the per-request price a first-class engineering constraint, not an afterthought.