Exa Launches Agent Ultra: A Subagent Swarm Deep Research API Built for Exhaustive List Building

Exa has pushed its Agent API into overdrive with a new mode called Agent Ultra. The pitch is blunt: this is the highest-effort setting, meant for research jobs that need to keep going until the well is dry — building huge lists, enriching entities, and chasing questions that pull in thousands of sources.

Under the hood, Exa says the system breaks one query into subtasks and sends out subagents across multiple domains at once. It uses frontier models where they matter and faster models where they’re enough. Ultra is the mode that spends the most compute, and Exa’s docs say it runs longer than any other effort level to return the most complete result it can.

The company says typical Ultra runs finish in about 30 minutes. Very hard tasks can stretch to 3 hours. That alone tells you what this product is really for: not casual search, but jobs where completeness matters more than speed.

Exa is backing the launch with benchmark claims across four tests: WANDR, DeepSearchQA, WideSearch, and Company Find-All. On its launch post, it says Ultra beat Opus 5.5, GPT-6 Astra, and Perplexity Agent at maximum effort on all four. The biggest gap shows up on Company Find-All, where Exa reports 2,451 passing entities per task, far ahead of the others listed.

But the company is also careful about the shape of those numbers. The comparisons are vendor-reported, not independently reproduced yet, and the benchmark setup is Exa’s own in some cases. Ultra is live now as a hosted API, not open weights, and it cannot be self-hosted. Pricing is metered at standard Agent usage rates, with runs capped by default at $20 unless the user sets maxCostDollars between $1 and $100. Exa also says you can stop a run early and keep the results, which is a nice way of saying the bill stops when the work does.