Exa unveils agent ultra: high-performance API for exhaustive research tasks

On September 26, 2026, Exa launched Agent Ultra, its most advanced version of the Exa Agent API, designed for complex research tasks requiring extensive data collection. The API excels in large list building, entity enrichment, and multi-source research, outperforming competitors like Opus 5.5, GPT-6 Astra, and Perplexity Agent on four benchmarks. Agent Ultra is available as a hosted API, with no open weights and no self-hosting option. It operates by splitting tasks into subtasks and assigning subagents to handle different domains simultaneously, utilizing frontier models where needed. Ultra mode consumes the most compute, running longer than other modes to deliver comprehensive results. Complex tasks typically complete in 30 minutes, while very hard tasks may take up to three hours. Exa reports a 9.1-point absolute gap over Opus 5.5 on the WANDR benchmark. The API supports outputSchema, input.data, and streaming, and users can test it via the Exa API Playground.