Exa’s September 25, 2026 launch post positions Agent Ultra as the run-to-exhaustion setting for web research: the mode you pick when a complete, cited list matters more than speed or cost.
What Exa published about Agent Ultra
Per Exa, Agent Ultra divides a task into subtasks and assigns subagents to research multiple domains at once, putting frontier intelligence where a task needs it and faster models where they are enough. Exa lists go-to-market uses such as building account lists (“every US company that makes browser-automation tooling”) and enriching a list with judgment fields, each with a cited URL as evidence — plus market maps for diligence, KYC research across filings and regulator sites, and training-data assembly. You can pass rows you already have so Ultra excludes them from new results.
“Today we shipped Exa Agent Ultra, setting a new state of the art for comprehensive deep research.”The Exa Team — Introducing Exa Agent Ultra, September 25, 2026
Budget, run time and benchmarks (from the primaries only)
The Agent Ultra documentation says Ultra is metered at the standard Agent usage rates up to a default $20 per run, and runs that finish early cost less. budget.maxCostDollars accepts $1 to $100; budget.maxDurationSeconds accepts 300 to 10,800 seconds. As a run nears either limit, the agent stops starting new work and returns what it has found. POST /agent/runs/{id}/stop ends a run early, keeps the results and bills usage up to the stop. The docs also note the SDK polling helpers time out after one hour by default, so long Ultra runs need a longer timeout or streaming. For a predictable price per request, Exa points users to a fixed effort instead.
| Line | Value | Range | Notes |
|---|---|---|---|
| Effort | ultra | — | Highest Agent effort |
| Cost cap | $20 default | $1–$100 | Standard Agent rates |
| Time cap | None default | 5 min–3 hr | Ultra-only control |
| WANDR | +12.6% vs Opus 5.5 | — | Exa-reported; half the cost per task |
Exa reports Ultra outperforming Opus 5.5, GPT-6 Astra and Perplexity Agent on their maximum effort settings across WANDR, DeepSearchQA, WideSearch and a Find-All Company test — for example +12.6% over Opus 5.5 on WANDR at half the cost per task. Attribute every benchmark and cost comparison to Exa’s post; this desk did not re-run them.
What Alberta operators should do
For Alberta sales and marketing teams, Ultra is a research engine for GTM research Alberta work: account lists (for example, Alberta oilfield-services firms that recently posted a safety leadership role), competitor maps, and enrichment fields with a source URL per cell. Start with one list you already build by hand, pass your existing rows to avoid duplicates, set maxCostDollars below the $20 default for the first runs, and spot-check the cited URLs before anything reaches a CRM or outreach sequence.
The same research discipline feeds AI search optimization — knowing which entities, competitors and sources answer engines cite — and the qualified-lead side of AI chatbot lead generation, where an enriched list tells your intake bot who is worth a fast human follow-up. For a quick primer on visibility in AI answers, see our AI search visibility Alberta guide, and keep reading the city desk at The Super Intelligence Times.
The guardrail
Ultra runs are long and metered, so budgets are the control — not an afterthought. Benchmarks are Exa’s own evaluations. Cited URLs still need a human check before outreach, and prospect lists touching personal data should respect Canadian privacy rules (PIPEDA / Alberta PIPA). A deep-research agent is not AGI, and a longer run is not the same thing as a verified answer.
effort: "ultra" on the Agent API, standard Agent rates up to a default $20 per run ($1–$100 settable), 5-minute to 3-hour time budgets, stop-early billing. Attribute WANDR and other benchmark wins to Exa only. No invented $. Pilot on one GTM list with a low cost cap and human URL spot-checks.