Exa Unveils Agent Ultra: AI Research Engine for Unprecedented Data Collection

Introducing Agent Ultra: The Next Generation of AI Research

Exa has announced the release of Agent Ultra, an enhanced version of its Exa Agent API aimed at tackling the most demanding research challenges. Positioned as the highest effort setting within the Exa Agent family, Ultra is specifically engineered for tasks requiring comprehensive data collection, extensive list compilation, entity enrichment, and analyzing information from thousands of diverse sources.

Internal testing by Exa indicates that Agent Ultra consistently surpasses other prominent AI systems, including Opus 5.5, GPT-6 Astra, and Perplexity Agent, when each is operating at its maximum capacity across four distinct research benchmarks.

Deployment and Functionality

Agent Ultra is immediately available as a hosted API service. Developers can access its capabilities by simply setting the 'effort' parameter to 'ultra' within the Exa API. It is important to note that this offering is not open-weights and does not support self-hosting, operating exclusively through Exa's cloud infrastructure.

At its core, Exa Agent Ultra operates by intelligently decomposing complex assignments into smaller, manageable subtasks. It then dispatches specialized subagents to concurrently investigate multiple information domains. The system strategically routes higher-capability frontier models to critical stages demanding advanced reasoning, while leveraging faster, more efficient models for less intensive steps. The 'Ultra' mode is optimized for maximum computational expenditure, guaranteeing extended operational times to deliver the most complete and thorough results possible.

Typical complex tasks processed by Ultra are often completed within approximately 30 minutes, though exceptionally challenging projects may require up to three hours to reach full completion.

Benchmark Performance Highlights

Exa's launch post details benchmark comparisons, with all competing systems evaluated at their peak performance settings. The reported figures, which are vendor-provided and have not yet undergone independent verification, showcase Agent Ultra's leading performance:

• WANDR (soft recall): Agent Ultra achieved 81.4%, significantly outperforming Opus 5.5 (72.3%), GPT-6 Astra (26.0%), and Perplexity Agent (40.1%).
• DeepSearchQA (F1): Ultra recorded 93.9%, topping Perplexity Agent (89.7%), GPT-6 Astra (85.3%), and Opus 5.5 (77.6%).
• WideSearch (row-level F1): With 58.9%, Ultra led against Perplexity Agent (56.0%), GPT-6 Astra (54.7%), and Opus 5.5 (51.6%).
• Company Find-All (average passing entities per task): Ultra identified 2,451 entities, a substantial increase over Opus 5.5 (146), GPT-6 Astra (113), and Perplexity Agent (98).

These performance advantages are reportedly accompanied by notable cost efficiencies. For example, Exa claims Ultra achieves a 12.6% improvement over Opus 5.5 on WANDR at half the cost per task, and a 4.7% lead over Perplexity on DeepSearchQA while being 46% more cost-effective than GPT-6 Astra.

Targeted Applications and API Details

Exa has identified several key user segments poised to benefit from Agent Ultra's capabilities:

• Model Providers: For assembling extensive training datasets, such as compiling every research paper or repository related to a specific technique, and verifying criteria like open-source weight releases.
• Financial Services: To construct detailed diligence market maps, conduct Know Your Customer (KYC) research across regulatory filings and court records, and continuously monitor portfolio signals.
• Go-to-Market Teams: For developing comprehensive account lists and enriching existing data with insightful judgment fields, each supported by a verifiable URL.

The system also supports expanding existing lists by excluding previously provided entries from new search results. Access to Ultra is facilitated via the standard Agent run endpoint, supporting features like output schemas, input data processing, and streaming results. Pricing is metered at standard Agent usage rates, with a default maximum cost of $20 per run. Users can adjust this budget from $1 to $100 and set maximum durations between 300 and 10,800 seconds. Runs can be stopped prematurely, retaining partial results and billing only for consumed resources.