Exa launches Agent Ultra, a high-compute deep-research API tier

Exa launched Agent Ultra, the highest effort tier of its Exa Agent API, aimed at research that must run to exhaustion: large list building, entity enrichment, and questions needing thousands of sources. Agent splits a task into subtasks, assigning subagents to research several domains at once and routing frontier models to steps that need them and faster models elsewhere; Ultra is the mode that spends the most compute. Exa says Ultra typically finishes complex tasks in about 30 minutes, with very hard tasks taking up to 3 hours.

Agent Ultra is live today on the Exa API by setting effort: 'ultra', and is also reachable through OpenAI compatibility mode on /responses. It is hosted-only: not open weights, not self-hostable. It uses the standard Agent run endpoint, supporting outputSchema, input.data, and streaming. Pricing is metered at standard Agent usage rates, up to a default $20 per run, with early finishes costing less. Budget controls include maxCostDollars ($1–$100) and maxDurationSeconds (300–10,800). A stop call ends a run early, keeps its results, and bills usage to that point; SDK polling helpers time out after 1 hour by default.

All figures come from Exa's launch post, with competitors run at their maximum effort setting, and none have been independently reproduced. Per Exa: WANDR soft recall 81.4% vs Opus 5.5 at 72.3%, GPT-6 Astra at 26.0%, and Perplexity Agent at 40.1%. DeepSearchQA F1 93.9% vs 77.6%, 85.3%, and 89.7%. WideSearch row-level F1 58.9% vs 51.6%, 54.7%, and 56.0%. Company Find-All averaged 2,451 passing entities per task vs 146, 113, and 98. Exa pairs each with a cost claim: on WANDR, a 12.6% gain over Opus 5.5 at half its cost per task; on DeepSearchQA, 4.7% over Perplexity at 46% lower cost per task than GPT-6 Astra; on WideSearch, 5.2% over Perplexity at the lowest cost per task of the four; on Company Find-All, 1,579% over Opus 5.5 at the lowest cost per entity. These gains are relative, not percentage points. On WANDR the absolute gap to Opus 5.5 is 9.1 points.

WANDR is Perplexity's benchmark of 500 wide and deep data-collection tasks with an open harness. Exa's grader shares the upstream evaluation logic but swaps in Exa as the contents tool, changes transport logic, and uses gpt-6-luna as the judge. Where a vendor 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.

Exa names three target groups. Model providers assembling training data — every paper and repo implementing a technique, verifying criteria like 'released weights, not just an API'. Financial services building diligence market maps, running KYC research across filings and court records, and monitoring portfolio signals. And go-to-market teams building account lists and enriching rows with judgment fields, each backed by a cited URL. Ultra can also expand an existing list: pass the rows you already have and they are excluded from new results. What Agent Ultra does not address is the gap between Exa's self-run benchmarks and independent reproduction, and the absence of open weights means no outside inspection of the system itself.