Exa 2.1 Delivers Sub-500ms Search with Frontier AI Accuracy

Search API provider Exa announces dramatic quality improvements across all endpoints, achieving both the fastest and most accurate search APIs through scaled pre-training and independent infrastructure.

NewDecoded

Published Nov 26, 2025

Nov 26, 2025

4 min read

Breaking the Speed-Quality Tradeoff

Exa released version 2.1 of its search API on November 23, 2025, delivering what the company calls a "dramatic improvement" in search quality across all product tiers. The release centers on two major updates: significantly enhanced quality for Exa Fast, Exa Auto, and Exa Deep endpoints, plus upgraded Model Context Protocol (MCP) integration with new deep search capabilities. Unlike competitors that wrap Google's search results, Exa built its own search engine from scratch, enabling performance characteristics impossible for wrapper-based services.

Speed Without Compromise

Exa Fast now operates at sub-500 millisecond latency while maintaining the highest accuracy in its category, according to the company's benchmarks. Most search APIs cannot break the 1,000ms barrier because they depend on Google's infrastructure with inherent speed limitations. Exa's independent document index and retrieval algorithms, developed over years of research, enable this speed advantage. The company ran evaluations using a SingleStep harness with GPT-4o-mini grading outputs against expected answers, completing tests on November 23, 2025.

Agentic Search for Maximum Accuracy

For applications prioritizing quality over speed, Exa Deep now represents the highest accuracy search API available. The Deep search type runs multiple searches sequentially to identify optimal results, accepting latency of a few seconds in exchange for superior relevance. In agentic evaluations completed November 20, 2025, Exa tested realistic scenarios where GPT-5 could invoke the MCP up to 10 times in sequence, demonstrating significant advantages over competing services.

Infrastructure at Petabyte Scale

Building a search engine required Exa to develop semantic and lexical databases from scratch, plus massive-scale crawling infrastructure spanning many petabytes. The company noted it has "rediscovered and then gone beyond retrieval techniques that only Google and Bing used to know." This infrastructure investment over multiple years forms the foundation for the quality improvements in version 2.1, which stem from scaling pre-training and test-time compute by an order of magnitude.

Continuous Scaling Ahead

Exa signaled that version 2.2 is already in development, with infrastructure scaling over recent months yielding consistent quality gains. The company is actively hiring engineers to build "one of the largest ML systems in the world," suggesting significant ongoing investment in the platform. The roadmap indicates quarterly or more frequent releases as training and index infrastructure continues to scale.

Decoded Take

Exa 2.1 addresses a fundamental constraint in AI application development: the forced choice between search speed and accuracy. By achieving sub-500ms latency with state-of-the-art accuracy through independent infrastructure, Exa is enabling real-time AI agents and RAG systems that were previously impractical for production deployment. This matters because most "AI search APIs" are merely wrappers around Google, offering no architectural differentiation or performance advantage.

As AI agents become more sophisticated and widespread, the infrastructure they depend on becomes critical bottleneck or enabler. Exa's willingness to invest years building search infrastructure from scratch, rather than taking the easier wrapper approach, positions it as genuine infrastructure for the emerging agent economy. The rapid release cadence (2.0 in October, 2.1 in November) and aggressive hiring suggest Exa sees search as a winner-take-most market where technical superiority compounds quickly.

Share this article