Octen launches AI-agent search with 10 ms results and a $20 plan

Search as part of the cost of running an AI agent

According to Octen, the company’s founder and CEO, Kuan Zou, previously led AI-powered search at Alibaba Cloud and enterprise search at Baidu. The startup names Square Peg and Argor among its investors, but has not disclosed how much it has raised. More than 5,000 developers have registered with Octen since the service’s general launch in late July.

Speaking to VentureBeat, Zou said that search speed and price limit how much information an AI agent can gather. In his view, machines can examine different sides of a question in parallel, making the cost of repeated queries especially important.

Zou also prefers to give the agent source materials to analyze itself, rather than first asking AI to produce a summary and then working from that. He said each stage can introduce errors.

For enterprise applications, this is a practical concern. A procurement assistant needs source data to distinguish a supplier’s current specifications from an outdated description. A coding agent needs relevant documentation, not just a plausible summary. The outcome depends on whether search can provide usable material quickly and cheaply enough to improve task performance.

What Pulse finds in 10 milliseconds

Octen’s research paper describes a generative search method. Instead of searching a conventional index and reranking pages, a trained model returns an identifier linked to an existing web address.

The 4-billion-parameter version of the model covers about 201 million pages. Its output is limited to identifiers in a table of web addresses, which is intended to prevent the language model from inventing URLs.

Octen’s published API specifications list median server-side latencies of about 10 milliseconds in lightweight mode and about 20 milliseconds in standard mode. Lightweight mode returns a page’s address and title; standard mode supports additional metadata, text snippets and, optionally, page content.

Search can quickly lead, for example, to a supplier’s status page, account recovery instructions or software documentation. But network transfer, loading additional content and the work of the model that prepares an answer each affect total response time.

The research results also show limits in coverage. Octen says the system ranks the relevant page first in 59.6% of navigational queries. A baseline dense retrieval method that takes clicks into account scored 46.5%. The paper’s authors also note that embedding-based search performs better on pages that are rarely visited. These are results from the company’s research, not proof that Pulse outperforms alternatives on every enterprise workload.

Pulse could be used as a fast way to find familiar URLs, supplemented by broader search when new sources are needed. But quickly finding a documentation home page does not prove that the service can find the specific passage an agent needs.

Subscriptions and enterprise costs

Octen’s pricing page lists a Standard subscription for $20 a month, including unlimited web search and 1,000 image searches per month. The $100 Plus subscription raises the image-search limit to 5,000 queries and gives requests priority over Standard.

Both plans are for personal accounts. During periods of heavy demand, requests may wait in a queue. Additional image search, fetching full page content and other APIs are charged separately, subject to the stated limits. A fixed subscription price therefore does not guarantee response times.

Published rates compare individual search queries, not completed information-gathering tasks:

These are different products and search modes, so this list does not show which service delivers results of comparable quality. The price competition is clear, however: Parallel’s fast modes cost the same as Octen search at its discounted rate.

For guaranteed throughput, Octen also sells resource-based plans while continuing to charge for API usage. The Builder plan costs $2,099 a month instead of $2,999 and provides for up to 50 requests per second and a service-level agreement. This is a different proposition from the $20 subscription.

When comparing offers, companies should consider not only the cost per query but also reserved throughput, parallel processing, content limits, support and data retention. Exa, for example, states that its enterprise offering retains no data and offers custom contract terms and service-level agreements.

Independent comparison of speed and price

Artificial Analysis’s search index, with data dated November 9, compares 26 products from 14 providers. The test keeps the answering model and verification settings the same while changing the search service.

In the highlights configuration, Octen scored 77. Exa auto and Perplexity medium scored 80 each, Exa lite scored 79, and Parallel advanced and Brave LLM context each scored 75.

Octen takes an average of 15.2 seconds per task. By comparison, Exa lite takes 19.0 seconds, Brave 22.0 and Parallel advanced 39.3. These times include measured search latency and estimated model runtime.

Octen search costs $9.07 per 1,000 tasks. The model adds another $14.50, bringing the total to $23.57. Exa lite’s total is $42.15 and Parallel advanced’s is $58.60.

The results support Octen’s advantages in speed and total cost per task, but show that some competitors score higher on quality. The test evaluated Octen Search, not the page search that Pulse claims can run in 10 milliseconds.

For an enterprise pilot, the most useful metric is the cost of an acceptable answer. A cheaper query loses its advantage if it leads to another search, misses important evidence or requires an employee to correct the answer.

Image search and licensing limits

Octen is adding image search. The company’s website says users can enter a text query or submit a reference image. Video search is still at an early stage. Octen suggests using found images to support visual content generation or including them alongside text in an agent’s response.

These features could be useful for comparing products and working with catalogs, but competitors already offer search across different media. Brave describes APIs for image and video search, while Parallel lists image-search pricing of $1 per 1,000 fast queries or $5 per 1,000 advanced queries. There is no evidence that Octen was the first to offer multimodal search.

For self-hosting, the license matters. The Octen-WebGR-4B model card lists CC BY-NC 4.0 for the model weights and the mapping table of identifiers to web addresses, which restricts commercial use. The inference code is distributed under Apache 2.0. The model card says to contact Octen for a commercial license.

Downloading the model therefore does not, by itself, give a company the right to use it in a commercial service. Buying API access and obtaining rights for local deployment are separate matters.

Verification instead of rationing queries

Octen’s central idea is to let AI agents check information more often without making every additional query an expensive system-design decision.

In an announcement, Zou described the company’s goal as making search available to developers as their workloads grow without requiring them to limit its use, much like a utility.

When evaluating the service, companies should test Pulse on tasks that require finding specific pages, compare general-search providers on typical research questions, and measure the quality of completed answers alongside latency and total costs.

Dataist’s take: search speed alone does not show how reliably an agent will complete a task. Production use depends on the quality of the material found, licensing terms and predictable performance under load.