What happened
In an October 28, 2026 blog post, Octen chief scientist Kunka introduced Pulse Search as a web-search API available at https://api.octen.ai/pulse-search. The service offers two modes. Lite returns a URL and title, with a stated median server-side latency of 10 ms and a price of $0.10 per 1,000 calls. Standard returns URL, title, favicon, author and date, and can optionally return relevant passages, full content or images. Its stated median latency is 20 ms and price is $0.20 per 1,000 calls. Full-content results cost $0.50 per 1,000 results, with one free result per call. Octen says these latency figures apply to queries the API has not seen before; repeated queries are faster.
The technical distinction is generative retrieval. Rather than retrieve candidates from a conventional index and rerank them, Octen says its Octen-WebGR-4B model generates a page identifier from a query. A lookup table converts that identifier into a URL. The company describes the identifier as two numbers: one representing a neighborhood of similar pages and another identifying a page within it. Because the model can only emit identifiers in the table, Octen says Pulse Search returns real pages rather than made-up URLs. The model was trained on aggregated click data without user identifiers; Octen says this teaches it which pages people seek for particular queries.
Octen claims WebGR-4B stores 200 million web pages in its parameters. It also describes WebGR-0.6B, a smaller model that memorizes 31.28 million pages and can run locally. In an internal demo on an M2 Pro MacBook, Octen reports 30–50 ms per lookup on-device, including a Chrome extension that generates URLs as a user types. The blog describes possible uses including type-ahead search, agent lookups, bulk matching of company names to official websites, and support systems that return help passages.
Octen says the WebGR-4B and WebGR-0.6B weights and page-ID tables are available on Hugging Face under CC BY-NC 4.0, with open-source inference code under Apache 2.0. The post links to a technical report, *Octen-WebGR: Scaling Generative Retrieval to 200 Million Web Pages*. It presents Pulse Search as generally available, but the supplied material does not specify a free tier, quota, service-level guarantee or geographic coverage.
What the archive adds
This is a new launch layered onto earlier coverage of the same model: Octen-WebGR-4B, a generative retrieval model for navigational search was already covered on November 2, 2026. The current announcement adds the API, latency and pricing claims, access details and the smaller on-device model. We also covered Expand Search, a web search API for agents, a separate provider's API.
Trend signal
The release supports the view that web search is becoming a composable capability for agent workflows: low-latency calls can be invoked more frequently than a conventional search pause might allow. This is adjacent to agent-ready software exposes capabilities, though the source establishes Octen's product design and claims, not wider adoption or a durable industry shift.
Why now
Octen's explanation is that conventional search latency interrupts model reasoning, especially when a task performs many lookups. The launch argues that a roughly token-scale wait makes always-on search and search-as-you-type practical. This is the company's stated rationale; the source does not independently measure the effect on end-to-end task performance.
Why it matters
If the advertised latency and price hold for real workloads, builders could stop treating web search as an expensive, separately rationed step. An agent could check more pages while reasoning, a command palette could search the web on each keystroke, and a support interface could fetch a relevant article before a user submits a ticket. The API also targets exact destinations such as official documentation, status pages and pricing pages, which are more useful than a generic list of results for many product tasks.
The model-and-table design is also a meaningful trade-off to evaluate. Restricting generated identifiers to known page IDs addresses fabricated URLs, while the retrieval model's memorized corpus may not have the freshness or breadth of a live index. Open weights and inference code create room for local experiments, but the non-commercial license on weights and tables limits some uses. Teams should test coverage, freshness, relevance and commercial rights against their own queries before making it a default dependency.
Limits and uncertainty
The latency, price, corpus size, accuracy framing and on-device measurements in the supplied material come from Octen. The source does not provide an independent benchmark, comparative relevance results, corpus update schedule, uptime data or details on handling pages outside the model's memorized set. A real-page constraint prevents invented URLs, but by itself does not establish that the returned page is current, authoritative or the best match. The post also does not establish that the claimed speed improves end-to-end agent quality or cost in production. The CC BY-NC 4.0 terms mean the model artifacts are not unrestricted for commercial use.
Reader takeaway
Pulse Search turns Octen's generative retrieval research into a priced API, while making model artifacts and inference code available under different licenses. Its headline is not simply faster search: it is the possibility of searching more often within a workflow. Validate retrieval quality, freshness and license fit on your own queries before relying on it.
Open question
How often are the page-ID table and memorized pages refreshed, and how does that affect results for fast-changing sites?