Exa Agent Ultra Ships for Exhaustive Web Research: Benchmarks, Cost and API Fit
Exa released Agent Ultra on September 24, 2026, adding a highest-effort mode to its research API for workloads where coverage and evidence collection matter more than latency or per-task cost. Exa positions Ultra for exhaustive list building, entity enrichment, due diligence and research tasks that can require thousands of sources.
The release builds on Exa Agent, which already decomposes large research jobs into subtasks and assigns subagents across multiple domains. Ultra raises the compute and search effort above the existing Agent tiers. Exa says its search index now covers more than 100 billion documents.
Exa also published benchmark results against OpenAI GPT-6 Astra, Anthropic Claude Opus 5.5 and Perplexity Agent. These are Exa-run or Exa-compiled evaluations that define the vendor's claimed performance envelope; independent production evaluation remains appropriate before selecting a research stack.
Agent Ultra benchmark results
Exa reports Ultra leading the compared systems on four research-oriented evaluations.
| Benchmark | Exa Agent Ultra | Best named comparison in Exa's table | Ultra median cost/task |
|---|---|---|---|
| WANDR | 81.4% | Claude Opus 5.5: 72.3% | $18.53 |
| WideSearch | 58.9% | Perplexity Agent: 56.0% | $3.85 |
| DeepSearchQA | 93.9% | GPT-6 Astra: 85.3% | Exa reports 52% lower cost than GPT-6 Astra |
| Find-All Company | 229 rows/task | GPT-6 Astra: 60 rows/task | GPT-6 Astra was cheaper on this benchmark |
For WANDR, Exa reports Ultra at 81.4%, compared with 72.3% for Opus 5.5, 40.2% for Perplexity Agent and 26.3% for GPT-6 Astra. The reported median cost per task was $18.53 for Ultra, $32.92 for Opus 5.5 and $23.09 for GPT-6 Astra.
WideSearch produced a narrower quality gap: 58.9% for Ultra, 56.0% for Perplexity Agent, 54.7% for GPT-6 Astra and 51.6% for Opus 5.5. Exa reports a $3.85 median cost per Ultra task in that evaluation.
On DeepSearchQA, Exa reports 93.9% for Ultra, 85.3% for GPT-6 Astra and 77.6% for Opus 5.5. On its internal Find-All Company benchmark, Ultra returned 229 successful rows per task versus 60 for GPT-6 Astra, 49 for Opus 5.5 and 48 for Perplexity Agent. Exa explicitly reports GPT-6 Astra as the lowest-cost system in that final comparison at $7.35 per task.
How the evaluation should be read
The benchmark provenance varies. Exa says its WANDR grader shares vendored evaluation logic with the upstream benchmark, with differences in the contents tool, transport logic and judge model. Where another provider had published a result on the same grader harness, Exa used that published figure; otherwise Exa ran the comparison itself.
That methodology makes the table a vendor evaluation with useful implementation detail. Production selection should still use a representative internal workload, especially when source coverage, citation quality, latency and cost have different weights from the public benchmarks.
What Ultra changes in the Exa Agent stack
Exa Agent was introduced in June 2026 as a research API that combines language models with Exa's retrieval stack. It decomposes complex requests into subtasks, can assign multiple subagents, and supports structured output for downstream applications.
The original Agent API exposed fixed effort levels from minimal through xhigh, with published per-request prices ranging from $0.012 for minimal to $1.00 for xhigh. Ultra is positioned above those modes for research where the stopping criterion is completeness instead of a small fixed search budget.
Exa describes the mode as using a mix of frontier and lower-cost models while parallel subagents investigate different parts of a task. The practical target is research that benefits from repeated discovery, qualification and evidence gathering beyond a single search-and-summarize pass.
Where Ultra is likely to fit
The strongest fit is wide-and-deep research: jobs that require finding many qualifying entities and then verifying multiple fields for every row. Examples supplied by Exa include market maps, KYC research, company enrichment, training-data discovery and searches for every public benchmark or implementation matching technical criteria.
That workload differs from latency-sensitive retrieval. A coding agent looking up one API signature or a chatbot grounding one answer usually benefits more from a fast search tier. Ultra spends substantially more time and compute to increase coverage, making it more appropriate for asynchronous research pipelines and high-value analyst workflows.
The output can also feed structured downstream systems. Existing Exa Agent supports an outputSchema parameter and caller-provided data, allowing research results to be constrained into application-defined records as well as prose.
Cost is workload-dependent
Ultra's launch post does not publish one universal flat price for every task. The benchmark data shows why: research depth and source count can change the amount of work substantially. Reported WANDR median cost was $18.53 per task, while WideSearch was $3.85 per task.
For comparison, the June Agent launch listed fixed prices of $0.012, $0.025, $0.10, $0.50 and $1.00 per request for the minimal, low, medium, high and xhigh effort tiers respectively. Teams should therefore route ordinary lookups and bounded research to lower effort levels and reserve Ultra for tasks where missing qualifying entities has a meaningful cost.
Deployment checklist
Before moving a production research workflow to Ultra, evaluate it on the same questions and schemas used in production. Track recall or completion rate, unsupported fields, source quality, duplicate entities, elapsed time and total task cost.
For list-building jobs, seed the agent with rows already collected when appropriate. Exa says Ultra can exclude supplied rows from future results, which is useful for incremental enrichment and recurring discovery jobs.
Applications handling regulated or sensitive research should separately validate their data-governance requirements and the suitability of every external source incorporated into the workflow. Exa includes KYC among its launch examples. Compliance decisions still depend on the application's own controls, approved data sources and review process.
Bottom line
Agent Ultra expands Exa from bounded search and research calls into a higher-compute mode aimed at exhaustive discovery. The most notable launch result is the combination of higher vendor-reported recall with explicit per-task cost data: Ultra leads Exa's four published comparisons, while the same tables show different cost leaders across workloads.
For developers, the decision is primarily a routing problem. Fast or fixed-cost Agent tiers remain better suited to routine retrieval. Ultra is designed for research jobs where incomplete coverage is expensive enough to justify a longer, more computationally intensive search process.
Sources
- Exa — Introducing Exa Agent Ultra – Lists without Limits (September 24, 2026): https://exa.ai/blog/exa-agent-ultra
- Exa — Introducing Exa Agent (June 16, 2026; architecture, effort tiers and published fixed pricing): https://exa.ai/blog/exa-agent
- MarkTechPost — Exa Launches Agent Ultra (September 26, 2026; independent current coverage): https://www.marktechpost.com/2026/09/26/exa-launches-agent-ultra-a-subagent-swarm-deep-research-api-built-for-exhaustive-list-building/