Editorial illustration for Exa's Agent Ultra API claims cost and performance gains over rivals
Exa's Agent Ultra API Cuts Costs, Beats Competitors
Exa's Agent Ultra API claims cost and performance gains over rivals
Exa turned on the highest-effort setting of its research API this week, calling it Agent Ultra. The company built it for jobs that don't have a clean stopping point: compiling exhaustive lists, enriching thousands of company or entity records, answering questions that require pulling from sources by the thousand rather than the dozen. It's live now on the Exa API, triggered by setting effort to "ultra," and it runs as a hosted service only. No open weights, no self-hosting.
Under the hood, Exa Agent breaks a research task into pieces and hands them to subagents working different angles at once, sending frontier models to the steps that need heavy reasoning and cheaper, faster models to the rest. Ultra is the setting that lets this process run longest. Exa's documentation puts typical completion time for complex tasks at around 30 minutes, with the hardest cases stretching to three hours.
The company is pairing the launch with benchmark claims against Anthropic's Opus 5.5, GPT-6 Astra, and Perplexity's Agent product, each tested at their own maximum effort setting, across four research-focused tests. The results, and what Exa says they cost per task, follow below.
Exa has released Agent Ultra, the highest effort level of its Exa Agent API. It is built for research that must run to exhaustion: large list building, entity enrichment, and questions that need thousands of sources. Exa team reports that Ultra beats Opus 5.5, GPT-6 Astra, and Perplexity Agent, each at maximum effort, on 4 research benchmarks.
Why this matters
For teams building on top of search and research APIs, Agent Ultra is worth testing against your own workloads, not just Exa's benchmarks. The numbers Exa published, +12.6% over Opus 5.5, +4.7% over Perplexity on DeepSearchQA, +5.2% on WideSearch, all come from Exa's own launch post, measured against competitors at their own maximum effort settings. That's a fine starting point, but it's Exa grading its own homework.
The cost claims are the more interesting hook here: half the per-task cost of Opus 5.5, 46% cheaper than GPT-6 Astra on one benchmark. If those hold up under independent testing, that's a real argument for anyone running large-scale entity enrichment or list-building pipelines where thousands of sources need to get touched. The catch is deployability: this is hosted-only, no open weights, no self-hosting.
So you're trading control for effort and cost gains, which is a fair trade for some teams and a dealbreaker for others handling sensitive data. Worth watching whether third parties replicate these benchmark gaps once Ultra sees broader use outside Exa's own test suite.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives AI Agent Internet Access Was 'Unintentionally Available,' Says Company CTO
From Archives Meta's Muse AI app fuels targeted ad push with user content
Common Questions Answered
What types of research tasks is Exa's Agent Ultra API designed to handle?
Agent Ultra is built for research jobs that require exhaustive completion without a clean stopping point, such as compiling comprehensive lists, enriching thousands of company or entity records, and answering questions that require pulling from thousands of sources rather than just a dozen. It's specifically engineered for large-scale research operations that need to run to exhaustion.
How does Agent Ultra compare to competitors like Opus 5.5 and GPT-6 Astra in performance benchmarks?
According to Exa's benchmarks, Agent Ultra outperforms Opus 5.5 by 12.6%, Perplexity Agent by 4.7% on DeepSearchQA, and achieves 5.2% gains on WideSearch benchmarks. These measurements were taken with all competitors running at their maximum effort settings, though the benchmarks come from Exa's own testing.
What deployment options are available for Agent Ultra?
Agent Ultra runs exclusively as a hosted service and does not offer open weights or self-hosting options. Users can access it through the Exa API by setting the effort parameter to 'ultra.'
Why should development teams test Agent Ultra against their own workloads?
While Exa's published benchmarks show performance improvements, teams building on search and research APIs should validate Agent Ultra's performance against their specific use cases rather than relying solely on Exa's internal benchmarks. This approach provides a more accurate assessment of whether the cost and performance claims translate to real-world benefits for their particular research needs.
