Exa launches Agent Ultra: maximum-effort mode for exhaustive research via API

Exa released Agent Ultra today, the highest-effort tier of its Agent API, built for research tasks that run to completion: building large lists, enriching entities, and answering questions that require thousands of sources. The company says Ultra beats Opus 5.5, GPT-6 Astra, and Perplexity Agent, each at its own maximum-effort setting, on four research benchmarks. The mode is available immediately through the API via the parameter `effort: "ultra"`. It is not open-weights and cannot be self-hosted.

How it works in practice

The system decomposes a task into subtasks and dispatches subagents that investigate multiple domains in parallel. It routes frontier models to stages that need them and faster models where they suffice. Typical Ultra runs finish in about half an hour; especially hard tasks can take up to three hours. Pricing is usage-based with a default ceiling of $20 per run; runs that finish earlier cost less.

The numbers the company cites

On Perplexity's WANDR benchmark — 500 broad and deep collection tasks — Ultra leads Opus 5.5 by 12.6% at half the cost per run; the absolute gap is 9.1 points. On Google DeepMind's DeepSearchQA (900 multi-step prompts) the advantage is 4.7% over Perplexity at 46% lower cost than GPT-6 Astra. On WideSearch, which tests broad collection, the lead is 5.2% over Perplexity at the lowest cost among the four. On Company Find-All, Exa reports a 1,579% advantage over Opus 5.5 at the lowest cost per entity found. All data comes from Exa's launch post; competitors ran at maximum effort; results have not been independently replicated.

Methodology and caveats

For WANDR, Exa used Perplexity's open evaluation harness, swapped the content tool for Exa, changed the routing logic, and ran gpt-6-luna as judge. Where a competitor had a published result on the same harness, Exa reported it; where not, Exa ran the evaluation itself. DeepSearchQA and WANDR tested up to 200 tasks per system; WideSearch and Company Find-All tested up to 100. The number of tasks actually graded varies by provider. In short, the figures are vendor-reported, not independently reproduced.

Target users and usage

Exa targets three groups: model providers building training data — for example, every paper and repo implementing a given technique, with criteria verification such as "weights released, not just API"; financial services doing market-mapping due diligence, KYC on filings and legal records, and portfolio signal monitoring; and go-to-market teams building account lists and enriching jurisdiction fields with cited URLs. Users can also expand an existing list by passing the rows they already have; those rows are excluded from the new results.

API, controls, and compatibility

The standard call supports outputSchema, input.data, and streaming. Users can set a cost ceiling of $1 to $100 (roughly 3.7 to 370 shekels) and a runtime of 300 to 10,800 seconds. A stop call ends the run early, preserves results, and charges only to that point. The SDK defaults to ending polling after one hour; users must extend the timeout or use streaming. OpenAI compatibility is also available: on the `/responses` path, set `reasoning.effort: "ultra"` with streaming or background mode. The Exa API Playground is available for testing.