Exhaustive research now has a higher setting. On September 26, 2026, Exa released Agent Ultra, the highest effort level of its Exa Agent API, built for research that must run to exhaustion. That means large list building, entity enrichment, and questions requiring thousands of sources — the sort of work that makes a normal search box look decorative.
Agent Ultra orchestrates parallel subagents and mixes frontier and faster models. Exa says the system beats Opus 5.5, GPT-6 Astra, and Perplexity Agent, with each competitor running at maximum effort, across four research benchmarks.
Four benchmarks, four different pressure tests
Exa reports an 81.4% soft-recall score on WANDR, a 93.9% F1 score on DeepSearchQA, a 58.9% row-level F1 score on WideSearch, and an average of 2,451 passing entities per task on Company Find-All. Those are the top scores Exa reports for all four tests.
WANDR is Perplexity’s benchmark of 500 wide and deep data-collection tasks, supported by an open harness. Agent Ultra scored 12.6% higher than Opus 5.5 and cost half as much per task; the absolute gap between the two systems was 9.1 points.
DeepSearchQA, created by Google DeepMind, contains 900 prompts for multi-step search. Ultra scored 4.7% higher than Perplexity and cost 46% less per task than GPT-6 Astra.
WideSearch tests broad information gathering. Ultra scored 5.2% higher than Perplexity and had the lowest cost per task among the four systems, while Company Find-All delivered a 1,579% improvement over Opus 5.5 with the lowest cost per entity found.
Exa evaluated up to 200 tasks each for WANDR and DeepSearchQA, plus 100 tasks each for WideSearch and Company Find-All. All figures come from Exa’s launch post, and the results remain vendor-reported rather than independently reproduced — a detail that belongs in the headline-sized facts, not the footnotes.
Built for long runs and structured output
Ultra runs typically finish complex tasks in about 30 minutes. Very hard tasks can take up to 3 hours, giving the service a maximum duration of 10,800 seconds; patience is now part of the API contract.
The service is available on the Exa API by setting effort: “ultra”. It uses the standard Agent run endpoint with outputSchema, input.data, and streaming support. On /responses, developers can set reasoning.effort: “ultra” with streaming or background mode.
Runs cost up to $20 by default, with the amount adjustable from $1 to $100. A stop call ends a run early and bills usage up to that point, while SDK polling helpers time out after 1 hour by default.
Ultra can also expand an existing list: developers pass existing rows, and the service excludes those rows from new results. That feature targets the practical mess behind list building, where finding more entities is only useful if the system does not keep returning the ones already found.
Exa lists model providers, financial services, and go-to-market teams as its three target user groups. The product is available through the API, but it is not open weights and cannot be self-hosted.
Agent Ultra is therefore less a new chatbot than a long-running research layer with structured output, parallel subagents, and a high ceiling for source gathering. Its benchmark lead is promising, but the claims still depend on Exa’s own evaluations; independent reproduction will decide whether the endurance is a breakthrough or just an expensive way to wait.
