News / AI & Data
Exa Agent Ultra: An Exhaustive Search API Deployed Only in Hosted Mode Published on 28 September 2026 by Christ-loisele (3 min read)
Exa has launched Agent Ultra, an ultra-performant version of its Exa Agent API, designed for in-depth searches and deployable exclusively as a hosted service. According to published benchmarks, it outperforms competitors on tasks requiring exhaustiveness and precision.
Video: American Ultra (2015) Official TV Spot - "New Kind of Agent" (Lionsgate Movies, YouTube) A Sub-Agent Architecture for Simultaneous Searches
Agent Ultra, the latest version of the Exa Agent API, stands out with its modular approach: it breaks down complex tasks into subtasks assigned to specialized sub-agents. These explore multiple domains simultaneously, thus optimizing source coverage. According to Exa, this method allows handling demanding queries in an average of 30 minutes, with a ceiling of 3 hours for the most challenging cases (source: MarkTechPost ).
Unlike solutions such as Opus 5.5 or GPT-6 Astra, Agent Ultra is not available in a self-hosted version or with open weights. Its deployment is limited to a hosted service, accessible via an effort parameter: 'ultra' in the API. This technical constraint reflects a priority: maximizing computational effort for exhaustive results, as highlighted in the official documentation (source: Exa.ai ).
Agent Ultra is designed for searches where exhaustiveness and precision matter more than speed or unit cost per task.
Illustration: Lawing Tech Performance Validated Through Targeted Benchmarks
Exa claims that Agent Ultra outperforms its main competitors: Opus 5.5, GPT-6 Astra, and Perplexity Agent, across four specialized benchmarks: WANDR (Wide And Deep Research), DeepSearchQA, WideSearch, and Find-All Company. For example, on WANDR, the model achieves a score of 81.4% while maintaining controlled costs (around 18 dollars per task), according to Eesel.ai . These results position Agent Ultra as a benchmark for tasks requiring exhaustive entity collection, such as compiling complete lists or data enrichment.
The WANDR benchmark, developed by Exa, specifically evaluates agents' ability to combine broad (wide) and deep research. Agent Ultra also excels in cost efficiency, a key criterion for businesses focused on profitability, as Exa emphasizes in its publications (source: Exa.ai ).
Targeted Use Cases: From Strategic Intelligence to Compliance
Agent Ultra targets sectors where precision and comprehensiveness take precedence over latency. Among the applications cited by Exa: building training datasets, verifying complex criteria (such as in due diligence), creating market maps, or monitoring dispersed signals across the web. For example, for a task like Find-All Company , the model generates an average of 2,451 validated entities per query, a volume unmatched according to benchmarks (source: Eesel.ai ).
Conversely, Exa notes that the tool is not suited for time-sensitive queries, such as API searches or anchoring chatbots. Its architecture, based on a mix of cutting-edge models (such as gpt-6-luna for evaluation) and lighter models for simple steps, also reduces token usage by up to 94 percent, thus optimizing costs (source: Eesel.ai ).
What this changes here
For Beninese businesses, government agencies, or organizations across West Africa, an API like Agent Ultra could transform processes currently managed manually or through less efficient tools. For instance, strategy consulting firms or financial institutions could use it for in-depth market analysis, reducing delays and errors tied to scattered research across multiple sources. Compliance services (such as anti-money laundering units) would also gain efficiency in systematically verifying entities, a critical issue in a context where cross-border flows are increasing.
However, its deployment, hosted exclusively, could raise data sovereignty challenges, particularly for regulated sectors. Variable costs depending on task complexity may also require careful modeling of needs before adoption. Finally, its focus on time-intensive tasks could limit its use for urgent needs, such as real-time competitive intelligence.
Sources 
