BlueOnyx
AIAI AgentsInfrastructureAPIArchitecturePerformance

10ms Web Search Enters the AI Agent Reasoning Loop

Théodore BaillyPublished on 14 novembre 20265 min read
Traînées de lumière sur une autoroute urbaine de nuit

Introduction

For years, web search has been the bottleneck in every AI agent pipeline: multi-second latency per query, pay-per-call costs that compound quickly as inference loops grow more complex, and response times fundamentally incompatible with real-time use cases. Octen, a San Francisco startup founded by Kuan Zou, formerly AI search lead at Alibaba Cloud and enterprise search researcher at Baidu, is looking to rewrite that equation with a retrieval model built for machines, not humans.

An Architecture That Breaks with Conventional Indexing

Octen's core technical proposition centers on what it calls generative retrieval. Where conventional search engines maintain an inverted index and rank results by estimated click probability, Octen trains a 4-billion-parameter model to directly generate identifiers pointing to relevant URLs. This architectural shift eliminates several intermediate processing steps, resulting in median server-side latency of 10 milliseconds in lightweight mode and 20 milliseconds in standard mode.

The model, called Octen Pulse, covers approximately 201 million web pages. On Artificial Analysis' Search API benchmark, Octen ranks as the fastest provider: 0.21 seconds per query and 15.1 seconds per complete task. On relevance, the model achieves a correct page placement rate of 59.6%, compared to a 46.5% baseline measured on the same benchmark.

Flat-Rate Pricing: A Structural Advantage for Development Teams

Beyond raw performance, it's the pricing model that's catching the attention of teams building agentic applications. The standard subscription at $20 per month unlocks unlimited web searches and 1,000 image searches. The Plus plan, at $100 per month, scales up to 5,000 image searches with priority access. For high-frequency production workloads, an Enterprise Builder tier is available at $2,099 per month for 50 requests per second.

This flat-rate logic stands in sharp contrast to the per-call pricing that dominates the competitive landscape. As agents make multiple search calls within a single reasoning session, an increasingly common pattern as pipelines grow more ambitious, unpredictable billing becomes a real barrier to scaling. A fixed monthly subscription fundamentally changes the unit economics of an agentic workflow.

A Segment in Active Consolidation

Octen isn't alone in this space. Exa, Tavily, Brave, and Parallel also offer agent-oriented search APIs, each making its own trade-offs around index coverage, latency, and metadata richness. What the emergence of this segment signals is a gradual specialization of the retrieval layer: teams that once wired up a generic consumer search engine are now treating this component as a first-class architectural decision.

Since reaching general availability in July 2026, Octen has onboarded more than 5,000 developers. The startup officially launched Octen Pulse and its updated pricing catalog on November 13, 2026, backed by Square Peg and Argor in a $10 million funding round.

For application owners and data teams managing RAG pipelines or decision-support assistants, the choice of search layer is now a concrete infrastructure decision, with direct implications for performance, cost predictability, and the ability to iterate confidently in production.

Share

10ms Web Search Enters the AI Agent Reasoning Loop