Exa launches Agent Ultra API for exhaustive research tasks
Exa has released Agent Ultra, the highest effort level of its Exa Agent API, designed for research tasks that require exhaustive coverage such as large list building and entity enrichment. The system splits tasks into subtasks and assigns parallel subagents, routing frontier models to complex steps and faster models to simpler ones. Exa reports that Agent Ultra outperforms Opus 5.5, GPT-6 Astra,…
Key points
- Exa released Agent Ultra, its highest effort API mode for exhaustive research and list building.
- Exa reports Agent Ultra beats Opus 5.5, GPT-6 Astra, and Perplexity Agent on four benchmarks.
- Runs cost up to $20 by default and typically take about 30 minutes to complete.
According to Exa’s launch post, Agent Ultra achieves a 12.6% improvement over Opus 5.5 on the WANDR benchmark at half the cost per task. It also shows a 1579% gain over Opus 5.5 on the Company Find-All benchmark. These figures are vendor-reported and have not yet been independently reproduced. Typical runs take about 30 minutes, though very hard tasks can take up to 3 hours.
Agent Ultra is available as a hosted API and is not open weights. Pricing is metered at standard Agent usage rates, with a default cap of $20 per run, adjustable between $1 and $100. The tool targets model providers, financial services, and go-to-market teams for tasks like training data assembly, due diligence, and account list enrichment.
Model pages: Opus 5.5 → · GPT-6 (Astra) → · GPT-6 Luna →
Exa Launches Agent Ultra: A Subagent Swarm Deep Research API Built for Exhaustive List Building
MarkTechPost · 26 September 2026
Loading the full article…
This text was published by MarkTechPost and written by Michal Sutter. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- OpenAI says AI agents leaked 53 images from ChatGPT users · 6 src
- Jeb model returns only judgments, up to 200 times faster · 1 src
- ChatGPT desktop app adds Appshots feature to read off-screen text · 1 src
- Google launches Intelligent Endpoints to make Chrome an AI execution platform · 1 src
- Oh My Pi updates Custom model configurations for version 18.2.7 and later · 1 src
Comments
via GitHub Discussions