Ex-Alibaba engineer's new startup Octen pairs 10-millisecond AI search with a $20 unlimited plan

A year-old startup called Octen is expanding its search service for AI agents with a model that locates web pages in roughly 10 milliseconds, a $20 monthly subscription for unlimited web searches, and image search.

For enterprises building assistants that repeatedly consult outside information, the proposition is straightforward: reduce the time and expense of looking things up so developers can let their systems check more sources.

The San Francisco company’s announcement combines Octen Pulse, its specialized page-finding model, with changes to its broader search offering. Its Pulse technical announcement, published October 28, describes the API as generally available.

But the headline numbers need context. Pulse’s 10 milliseconds measures median processing time on the server, not the entire journey from an enterprise application to a finished answer.

The unlimited plan covers Web Search and can queue requests. And the downloadable Pulse model has noncommercial restrictions, a consequential distinction from the release’s open-source characterization.

Those qualifications do not erase the opportunity. They define where Octen could fit alongside competitors including Exa, Parallel, Brave and Tavily—and what buyers should test before committing.

Search becomes part of the agent’s operating budget

Founder and CEO Kuan Zou previously led AI search at Alibaba Cloud and enterprise search at Baidu, according to Octen.

The company names Square Peg and Argor as investors without disclosing a funding amount in the announcement, and says more than 5,000 developers have registered since its broader service became generally available in late July.

In an interview with VentureBeat, Zou described search speed and cost as constraints on how much information an agent can gather. His argument is that machines can investigate several aspects of a question concurrently, making the economics of repeated searches especially important.

“We have to reduce the cost,” Zou said.

He also favors giving an agent source material it can interpret itself. “If you use AI to generate the summary and then use AI to consolidate the summary, sometimes you make some mistakes, right?” he said.

That distinction matters for enterprise applications. A procurement assistant comparing suppliers needs enough underlying evidence to distinguish a current specification from an outdated description. A coding agent needs the relevant documentation, not simply a plausible summary. The practical question is whether a search service supplies usable evidence at a cost and speed that improve the completed task.

What Pulse retrieves in 10 milliseconds

Octen’s research paper describes a technique called generative retrieval. Instead of searching a conventional index and repeatedly ranking candidate pages, a trained model produces an identifier that maps to an existing URL.

The 4-billion-parameter version covers approximately 201 million pages. Its output is constrained to identifiers in a URL table, addressing the problem of a language model inventing a web address.

The published API specifications put lite mode at about 10 milliseconds and standard mode at about 20 milliseconds, both server-side medians. Lite supplies a URL and title; standard supports additional metadata and optional passages or page content.

For a developer, the clearest application is quickly finding a destination: a vendor’s status page, an account recovery guide or a software documentation page. Network travel, fetching additional content and the answering model’s work remain separate contributors to response time.

The research also reveals a coverage trade-off. Octen reports 59.6% accuracy for placing the correct page first across its navigational traffic, compared with 46.5% for a click-enhanced dense retrieval baseline. Its paper finds that embedding-based retrieval retains an advantage on infrequently visited pages. These are company-reported research results, not proof of superior performance on every enterprise workload.

The architectural implication is to evaluate Pulse as a fast route to familiar destinations, with broader retrieval available when the question requires discovery. A tool that rapidly finds the right documentation homepage still needs to prove it can supply the exact passage an agent needs.

A low monthly price, with different terms for production

Octen’s pricing page lists a $20 Standard subscription with unlimited Web Search and 1,000 image searches monthly. A $100 Plus subscription raises the image allowance to 5,000 and receives priority over Standard.

Both are personal-account plans, and requests can queue during busy periods. Additional image searches, full-content retrieval and other APIs carry separate usage charges, subject to listed allowances. Enterprises should therefore distinguish predictable spending from guaranteed response time.

The following published rates compare search requests, rather than complete research tasks:

These are different products and retrieval modes, so the table does not establish an equivalent-quality winner. It does show that Octen faces price competition: Parallel’s fast modes match its discounted general-search rate.

For guaranteed throughput, Octen separately sells capacity plans while continuing to bill API usage. Its Builder tier lists $2,099 monthly, discounted from $2,999, for up to 50 queries per second and a service-level agreement. That is a substantially different purchasing decision from the $20 subscription.

Buyers should compare the commercial terms around each rate: reserved capacity, concurrency, content allowances, support and retention. Exa, for example, lists zero data retention for its enterprise offering, with custom contractual and service-level options.

Independent results show a speed-cost trade-off

The Artificial Analysis Search Index, with data dated November 9, evaluates 26 products across 14 providers. It holds the answering model and testing setup constant while changing the search provider.

Octen’s highlights configuration scores 77, compared with 80 for Exa auto and Perplexity medium, 79 for Exa lite, and 75 for Parallel advanced and Brave’s LLM context configuration.

Octen records 15.2 seconds per task, versus 19.0 seconds for Exa lite, 22.0 for Brave and 39.3 for Parallel advanced. The timing combines measured search time with derived model time.

Its search component costs $9.07 per 1,000 tasks, with $14.50 in model costs, totaling $23.57. Exa lite totals $42.15; Parallel advanced totals $58.60.

Those results support Octen’s case on speed and total task expense while showing that some competitors achieve higher quality scores. They evaluate Octen Search, however, not Pulse’s 10-millisecond retrieval claim.

For an enterprise pilot, the useful measurement is cost per acceptable answer. A cheaper call loses its advantage if it forces repeated searches, misses important evidence or produces an answer employees must correct.

Images broaden the offering; licensing narrows self-hosting

Octen also adds visual retrieval. Its product site describes image searches using either text or a reference image, while video search remains in early access. The company suggests using retrieved images to support visual generation or combine pictures with text in an agent’s response.

This could help product-comparison and catalog workflows, but access to multiple media types is already competitive territory. Brave documents image and video endpoints, and Parallel publishes image-search pricing of $1 per 1,000 fast requests or $5 for advanced requests. The evidence does not establish Octen as the first multimodal search provider.

Self-hosting requires another distinction. The Octen-WebGR-4B model card specifies CC BY-NC 4.0 for the weights and identifier-to-URL table, restricting commercial use. Its inference code uses Apache 2.0. The card directs commercial users to contact Octen for licensing.

Consequently, a business cannot assume that downloading the model gives it permission to deploy a revenue-generating service. API procurement and commercial rights for local deployment are separate questions.

The enterprise opportunity is more frequent verification

Octen’s strongest pitch is that agents should be able to verify information more frequently without making each additional check an expensive design decision.

“We’re building search the way utilities deliver water and electricity, so developers never have to ration it as they grow,” Zou said in the announcement.

The evidence suggests a practical evaluation path: test Pulse on destination-finding tasks, compare general search providers on representative research questions, and measure completed-answer quality alongside latency and total spending.

For technical leaders, the decision ultimately turns on whether faster retrieval changes what their agents can reliably accomplish. Finding a page in milliseconds is valuable. Finding the right evidence, with appropriate commercial terms and dependable production performance, is what makes it useful infrastructure.