Okay, I finally did it. I've seen the "fastest AI search" claims, and as someone who lives in docs and needs quick answers, I had to test it. I ran 500 complex, multi-part queries over a week—things like "compare NextJS 13 app router vs pages router for SEO in edge deployments."
The reality? It's *fast*, but with a big caveat. The initial response is blazing, often under 2 seconds. But for truly complex queries, the "full" answer with all sources loaded and the final, refined output? That often added 5-8 seconds of waiting. The speed claim feels true for the first byte, not necessarily the complete, usable answer.
The pattern I noticed: simpler factual searches are incredibly snappy. But the moment you ask for synthesis or a detailed comparison, you hit a "loading sources" wall. For my workflow, that initial speed is great for quick lookups, but the total time to a thorough answer isn't that different from other tools.
Still, the sheer consistency was impressive—no major slowdowns during peak hours. The speed is real, just maybe not exactly as you'd interpret it. It's about that first burst, not always the finished product.
I'm a data engineering lead at a mid-market ecommerce company. We run analytical workloads on ClickHouse for real-time dashboards and have tested several OLAP engines for ad-hoc querying.
Core comparison for "fastest AI search" type tools:
1. Target audience - You're looking at individual power users or small teams. The pricing and feature set aren't built for enterprise-grade SLAs or complex governance.
2. Real pricing - Expect $10-20/user/month for the pro tiers that unlock the higher rate limits necessary for heavy use. The hidden cost is the time spent waiting for full source synthesis on complex queries, as you noted.
3. Where it breaks - The architecture is optimized for retrieval of initial context, not for deep synthesis of multiple complex documents in real-time. Your benchmark matches what I'd expect: sub-2s for initial response, 5-8s+ for full synthesis. That's a hard trade-off.
4. Where it wins - Unmatched consistency for straightforward factual lookup. In our internal tests, response time variance was under 400ms for simple queries even under load, which is impressive.
My pick: I'd recommend sticking with it for quick lookups and initial research, but pair it with a tool for deeper analysis. If your primary use case is simple factual retrieval, it's the best. If you need consistent under-5s full synthesis for complex comparisons, you'll be disappointed. Tell us the ratio of simple to complex queries in your workflow and your max acceptable wait time for a complete answer.
Numbers don't lie.