Octen - We stayed quiet and built the FASTEST AI search on Earth
by•
Octen is a search API built for AI agents, not humans. Instead of one query at a time, it decomposes a question into sub-queries, fires them all at once, and returns results in 62ms (P50), several times faster than other providers. Pricing is $1 per 1,000 calls, well below typical rates of $4 to $9. A single account handles over 1,000,000 queries per second, with new content indexed within 5 minutes. Our embedding models rank #1 on RTEB and MMEB-v2.
Replies
Hey Product Hunt 👋
I'm Kuan, founder of Octen. Before this, I led AI search product at Alibaba Cloud and built Baidu's enterprise search platform, so I've spent a decade dealing with search infrastructure that had to survive real scale.
When I looked at the search APIs being handed to AI agents today, I was surprised. Most are the same human-search stack wrapped in an endpoint, and many start to break once you push tens of queries per second. But an agent doesn't search the way a person does. It wants to ask a question from many angles at once and get everything back together.
So we rebuilt search from scratch for that use case. Octen decomposes a query into sub-queries, fires them concurrently, and returns results in 62ms (P50) and 68ms (P90) on SealQA Hard, several times faster than other providers in this space, at $1 per 1,000 calls instead of the usual $4 to $9.
We spent the past several months building this quietly because we didn't want to make noise before the numbers were real. All benchmarks are public and reproducible (link in the post). Would love for you to run it against whatever you're using today and tell us what you find.
Happy to answer any questions here.
Honestly the sub-query thing is kind of clever, I tested it on a research prompt and it came back faster than I expected. Price point looks solid too.
The 5-minute indexing window is a really smart call for agent workflows where freshness actually matters. Love that the whole thing is tuned for machine consumers instead of bolting an API onto a human-facing search UI.
Love that the 62ms P50 is broken down with the sub-query decomposition approach shown so clearly, that level of transparency on latency methodology is rare. Pricing at a dollar per thousand while claiming that speed is the kind of detail that makes me want to actually test it myself.
Currently as an engineer building at Octen, I excited for our launch. I have been applying our search capabilities to wearables and AR, and that's where latency actually bites. When a device is reacting to what you're looking at, search that takes a second breaks the whole thing. Sub-100ms is what makes it feel like it just knows.