Exa adds a highest-effort mode that splits work across subagents

single source· 1 articles · confidence: medium · first seen 2026-09-26 08:04 UTC

What this means for you

Nothing to act on yet unless list building is your bottleneck. Exa quotes no price, rate limit or latency, and the benchmark table is its own, undated work — so it cannot be set beside anyone else's run. Treat the claims as vendor statements until an outside party reruns them.

Exa has released Agent Ultra, the highest-effort setting of its Exa Agent API, which splits a request across multiple subagents — parallel workers, each searching separately — to compile lists and enrich entity records from thousands of sources. Exa says it outperforms Opus 5.5, GPT-6 Astra and Perplexity Agent on four benchmarks, including 81.4% soft recall on WANDR (a scoring rule that credits close matches, not only exact ones). Every figure is Exa's own. The announcement gives no evaluation date, harness or per-benchmark breakdown, and no pricing or rate limits.

Models in this story

Key facts

  • ·Exa released Agent Ultra as the highest-effort mode of its Exa Agent API, published 26 September 2026 source
  • ·The mode coordinates subagents across thousands of sources for list building and entity enrichment source
  • ·Exa reports it beats Opus 5.5, GPT-6 Astra and Perplexity Agent on four benchmarks source
  • ·Exa reports 81.4% soft recall on WANDR source
  • ·No evaluation date, harness, pricing or rate limit is given in the announcement source

What the sources say

  • MarkTechPost — Vendor-side write-up of the release that carries Exa's benchmark numbers without saying how they were run

Sources

The original reporting. Follow these — they did the work.

← the wire