Exa Agent Ultra is the top "effort" tier of the Agent API that AI search company Exa released on September 26, 2026: set the effort parameter to "ultra" and the agent throws maximum compute at tasks that must be run to exhaustion — large-scale list building, entity enrichment, and problems that require digging through thousands of sources. It doesn't aim to be faster; it aims to be exhaustive.
From one question to thousands of sources: how it runs
Ultra works by splitting a task into subtasks, then orchestrating parallel subagents that research multiple domains at once; each step mixes frontier models with faster, cheaper ones as needed. It is Exa's most compute-intensive, most time-consuming mode: a typical complex task finishes in about 30 minutes, while extremely hard ones can run up to 3 hours (maxDurationSeconds ranges from 300 to 10800). It's already live: set effort="ultra" in the Exa API, and it also works through OpenAI's /responses interface (reasoning.effort="ultra"). Note that these are not open weights — you can't self-host it.
Pricing and who it's for: a cost switch from $1 to $100
Billing follows standard Agent usage metering: the default per-run cap is $20, adjustable from $1 to $100 (maxCostDollars); runs that finish early are billed on actual usage only. SDK polling times out after one hour by default, and the official guidance is to extend it or switch to streaming. The three target audiences Exa names are: model builders (assembling training data, e.g. collecting every paper and repo on a technology), finance teams (diligence market maps, KYC, portfolio monitoring), and go-to-market teams (customer list building and enrichment, each result carrying its source URL). You can also pass in an existing list for expansion — the rows you pass in are automatically excluded.
The benchmarks look strong — but they're all vendor-reported
The official scorecard is genuinely impressive: on WANDR (Perplexity's 500 wide-and-deep data collection tasks), 81.4% soft recall, 12.6% above Opus 5.5 in relative terms (a 9.1-point absolute gap), at half the per-task cost of Opus 5.5; on DeepSearchQA (Google DeepMind's 900 multi-step search prompts), 93.9% F1, 4.7% above Perplexity; on WideSearch, 58.9% row-level F1, 5.2% above Perplexity, with the lowest per-task cost of the four systems; on Company Find-All, an average of 2,451 passing entities per task, about 16.8x the 146 from Opus 5.5. But the sourcing must be stated plainly: all of the above comes from Exa's own release, every competitor was run at its maximum effort setting, and none of these numbers has been independently reproduced yet.
"Exhaustive" finally has a price tag
Deep research is shifting from "give me an answer" to "enumerate exhaustively", and Ultra's value is in turning "thoroughness" into an API parameter you can switch on against a budget — a cost cap of $1 to $100. That is exactly what makes it enterprise-usable: the performance ceiling and the spending ceiling can both be pinned down. But until third parties reproduce the results, treat this scorecard as a reference line for the capability ceiling, not as a basis for procurement decisions.
