A parallel harness stretches the search

Ultra uses a custom research harness that breaks a query into subtasks, assigns them to multiple model workers, searches several domains in parallel, and merges the findings. The system relies on three main components:

• Subagent orchestration for parallel searches across sources and subject areas
• Code execution for filtering, deduplication, validation, and structured output assembly
• Content highlights that extract relevant passages from pages before sending material to a model

Exa attributes much of Ultra’s efficiency to the highlights system. Full web pages can consume large context windows, increasing inference cost and latency. Exa says its highlights model has reduced retrieved-page token use by as much as 94% in testing. The 94% figure concerns the highlights model; end-to-end savings vary by run.

Dynamic Highlights extends that approach by selecting excerpts across the entire candidate result set. This gives the agent a smaller set of relevant passages when it compares many sources supporting or disputing the same claim.

Strong scores, bounded evidence

Exa reports that Ultra leads four evaluations: WANDR, WideSearch, DeepSearchQA, and an internal Find-All Company benchmark. The announcement provides comparative figures for the first three:

On WANDR, Exa reports that Ultra costs 44% less per task than Opus 5.5 and 20% less than GPT-6 Astra. A WideSearch task costs $3.85, according to Exa, which is 25% below Perplexity Agent, 40% below Opus 5.5, and 54% below GPT-6 Astra. On DeepSearchQA, Ultra leads GPT-6 Astra by 8.6 percentage points and Opus 5.5 by 16.3 points.

WANDR, short for Wide And Deep Research, measures how well an agent finds a large set of qualifying entities and supplies evidence across multiple fields. Its tasks resemble company research, due diligence, and legal discovery, where omitted entities reduce the value of the result.

Cross-vendor benchmark results depend heavily on the evaluation harness, tool access, effort limits, and judge model. Exa says its WANDR grader uses evaluation logic copied from the upstream repository, with differences in the content tool, transport layer, and model that scores answers. It also uses competitors’ published figures when results from the same harness are available. Parallel maintains a separate DeepSearchQA leaderboard with different results, illustrating how implementation choices affect rankings. These scores measure performance on benchmark tasks and cannot guarantee complete coverage of the open web.

Put Ultra behind a job queue

Agent Ultra is available through the API documentation. Existing Exa Agent integrations can select it by setting effort to "ultra" and supplying optional cost and duration limits:

import Exa from "exa-js";

const exa = new Exa();

const run = await exa.agent.runs.create({
  query: "Find all companies building browser automation tools in the United States.",
  effort: "ultra",
  budget: {
    maxCostDollars: 10,
    maxDurationSeconds: 1800
  }
});

const finished = await exa.agent.runs.pollUntilFinished(run.id, {
  timeoutMs: 3 * 60 * 60 * 1000
});

The client creates an asynchronous run and then polls for completion. In this example, the server-side research budget is 30 minutes, while the client allows polling to continue for as long as three hours.

• Ultra uses standard Agent metered billing and has a default spending cap of $20 per run.
• budget.maxCostDollars changes the spending cap.
• budget.maxDurationSeconds accepts durations from five minutes to three hours.
• When the duration expires, the run returns the results collected so far.
• Exa says complex Ultra runs typically finish in about 30 minutes, while difficult tasks can take up to three hours.
• An HTTP POST to /agent/runs/{id}/stop ends a run early and preserves its collected results.
• Duration limits and early stopping require effort: "ultra". Cost limits also work with effort: "auto".

Batch research is the natural fit

Ultra suits workloads where broader coverage justifies higher latency and spending. Common applications include:

• Building lists of companies or people that meet difficult-to-verify criteria
• Running competitive research, sourcing, and sales-prospecting jobs
• Supporting due diligence or legal research where omitted entities create material gaps
• Enriching seed lists with structured fields gathered from scattered web sources
• Answering multi-hop questions that require evidence from several domains

Latency-sensitive agent loops, inexpensive retrieval-augmented generation, and interactive chat fit the automatic or lower-effort modes more closely. A research run that lasts 30 minutes and can spend up to its configured cap belongs in an asynchronous workflow with status tracking, cancellation, and persistent results.

Exa bets on leaner context

Deep-research products from Exa, Parallel, Perplexity, OpenAI, Anthropic, and Google share a broad architecture: a model receives search tools, a budget, and time to investigate. Their performance depends on how they choose queries, divide work, retrieve sources, manage context, verify evidence, and stop.

Exa’s strategy centers on a specialized search index and passage extraction that reduces the amount of page content sent to language models. Its published results support that approach for wide entity searches and multi-step research, subject to the limits of vendor-run evaluations.

Teams evaluating Ultra should test representative queries and measure entity recall, factual precision, citation quality, wall-clock time, and total cost. Those workload-specific results will provide a stronger deployment signal than cross-vendor benchmark rankings alone.