Exa launches Agent Ultra, a subagent-swarm research API for exhaustive list building
Exa has released Agent Ultra, the highest effort level of its Agent API, which orchestrates parallel subagents for large-scale list building and entity enrichment. Exa reports top scores against Opus 5.5, GPT-6 Astra and Perplexity Agent on four benchmarks, with runs metered at up to $20 by default.
Exa Agent splits a task into subtasks and assigns subagents to research several domains at once, routing frontier models to the steps that need them and faster models where those are enough. Ultra is the mode that spends the most compute. According to the Agent Ultra documentation, it runs longer than any other effort level to return the most complete results. Typical complex runs finish in about 30 minutes, while very hard tasks can take up to three hours.
Exa reports that Ultra beats Opus 5.5, GPT-6 Astra and Perplexity Agent, each run at its maximum effort setting, on four benchmarks. On WANDR, measured by soft recall, Agent Ultra scored 81.4%, against 72.3% for Opus 5.5, 26.0% for GPT-6 Astra and 40.1% for Perplexity Agent. On DeepSearchQA, measured by F1, Ultra scored 93.9%, against 77.6%, 85.3% and 89.7% respectively. On WideSearch, measured by row-level F1, Ultra scored 58.9%, against 51.6%, 54.7% and 56.0%. On Company Find-All, measured by average passing entities per task, Ultra returned 2,451, against 146 for Opus 5.5, 11 for GPT-6 Astra and 398 for Perplexity Agent.
Exa pairs each result with a cost claim: on WANDR it reports a 12.6% gain over Opus 5.5 at half that system's cost per task; on DeepSearchQA a 4.7% gain over Perplexity at 46% lower cost per task than GPT-6 Astra; on WideSearch a 5.2% gain over Perplexity at the lowest cost per task of the four systems; and on Company Find-All a 1579% gain over Opus 5.5 at the lowest cost per entity found. Exa describes those figures as relative gains, not percentage points, and notes that the absolute gap to Opus 5.5 on WANDR is 9.1 points.
On methodology, WANDR is Perplexity's benchmark of 500 wide and deep data-collection tasks with an open harness. Exa says its grader shares the upstream evaluation logic, swapping in Exa as the contents tool, changing transport logic and using gpt-6-luna as the judge. Where a vendor had published a result on that harness, Exa reports that figure; otherwise Exa ran the benchmark itself. DeepSearchQA is Google DeepMind's 900-prompt multi-step search benchmark, and WideSearch tests broad information gathering. Exa evaluated up to 200 tasks each for WANDR and DeepSearchQA and 100 each for WideSearch and Company Find-All, with graded task counts varying by provider. All results are vendor-reported and have not yet been independently reproduced.
Exa lists three target user groups. Model providers can assemble training data, such as every paper and repository implementing a given technique, and verify criteria such as released weights rather than just an API. Financial services firms can build diligence market maps, run KYC research across filings and court records, and monitor portfolio signals. Go-to-market teams can build account lists and enrich rows with judgment fields, each backed by a cited URL. Ultra can also expand an existing list: rows passed into the request are excluded from new results.
Ultra uses the standard Agent run endpoint, and requests support outputSchema, input.data and streaming. Pricing is metered at standard Agent usage rates, up to a default of $20 per run, with runs that finish early costing less. The maxCostDollars parameter accepts $1 to $100, and maxDurationSeconds accepts 300 to 10,800 seconds. A stop call ends a run early, keeps its results and bills usage up to that point. SDK polling helpers time out after one hour by default, so Exa advises setting a longer timeout or streaming events. For OpenAI compatibility, users can set reasoning.effort to "ultra" on the /responses endpoint with streaming or background mode, and test the model in the Exa API Playground.