Exa has introduced Agent Ultra, a new API mode designed for deep research and exhaustive list building. The tool operates as a subagent swarm, breaking down complex tasks into smaller sub-assignments that are processed concurrently across multiple domains. It dynamically routes tasks between frontier models and faster, more efficient models based on the complexity of the specific step.
The system is available as a hosted API and does not support self-hosting or open weights. Typical task completion takes approximately 30 minutes, though highly complex requests may require up to 3 hours. Users can integrate the tool using the standard Exa Agent run endpoint, which supports streaming, output schemas, and input data processing.
Pricing is metered based on standard Agent usage rates, with a default maximum cost of $20 per run. Users can configure budget limits between $1 and $100 per request. The system also allows for duration constraints ranging from 300 to 10,800 seconds. If a run is stopped early, the service bills only for the usage incurred up to that point.
Exa evaluated the performance of Agent Ultra using benchmarks such as WANDR, DeepSearchQA, and WideSearch. The company reports that on the WANDR benchmark, the absolute gap to Opus 5.5 is 9.1 points. These results are based on vendor-reported data and have not yet been independently verified by third parties.
