Home Docs Blog
Benchmark Book a Demo Login

The new Sixtyfour search, tested on Exa's own benchmarks

Sixtyfour search now puts a match first on 90.4% of the queries in Exa's people search benchmark, ahead of every published result.

Yahia Salman Sixtyfour Research
October 13, 2026 3 min read Product
SearchPeople SearchCompany SearchBenchmarksSearch APIProduct Updates
The new Sixtyfour search, tested on Exa's own benchmarks

We rebuilt how Sixtyfour search finds people and companies. On Exa's open people search benchmark, our first result now matches the search on 90.4% of queries. Exa publishes 72.0% for its own search.

Bar charts of precision and R@1 on Exa's people and company search benchmarks. Sixtyfour leads every chart: people precision 82.8%, people R@1 90.4%, company precision 79.6% and company R@1 86.4%.
Sixtyfour against the results Exa publishes for Exa, Brave and Parallel, and the figure Juicebox reported.

What changed

You can describe the people or companies you want in plain English, the way you would brief a colleague. Several agents work on that request together, each writing part of the filters the search runs on.

A reranking model then reads the results and judges how well each one fits what you asked for. The strongest fits move to the top of your list, and results that clearly don't fit are dropped. You spend less time skipping past people in the wrong role or the wrong city.

This applies to natural-language people and company searches in the Sixtyfour app, the API and the MCP server.

The results

Exa released its people search and company search benchmarks with the queries, the grading prompts and an evaluation harness in an open source repository, so anyone can run them against their own search. The people benchmark has 1,400 queries like "director of marketing based in california senior level", and the company benchmark's retrieval track has 605. An LLM judge reads each result and decides whether it fits the query.

Table of Sixtyfour against Exa on all six measures, with Sixtyfour's lead in points. People search: R@1 90.4% against 72.0%, R@10 96.6% against 94.5%, precision 82.8% against 63.3%. Company search: R@1 86.4% against 61.8%, R@10 94.4% against 94.2%, precision 79.6% against 65.9%.
Sixtyfour against the results Exa publishes for its own search, on every measure of both benchmarks.

Sixtyfour reaches 82.8% precision on the people search benchmark, ahead of Juicebox (79%), Exa (63.3%), Brave (30.2%) and Parallel (26.9%).* Precision is the average share of returned results that are relevant. R@1 measures how often the top result is relevant, and R@10 how often at least one relevant result appears in the top ten.

We graded with Exa's prompts unchanged, using GPT-5.4, the default in Exa's benchmark scripts until September 2026. You can try the new search in the Sixtyfour app or through the API, and the queries and grading are public if you want to check these numbers yourself.

*Exa, Brave and Parallel: the results Exa publishes in its repository. Juicebox: its own reported figure.

Frequently asked

An open benchmark Exa published with 1,400 role-based queries, such as a job title in a particular city. An LLM judge reads each result and decides whether the person holds that role in that place. Exa also published a company search benchmark, whose retrieval track has 605 queries.

R@1 is the share of queries where the first result matches the search. R@10 is the share where a match appears in the top ten results. Precision is the average share of returned results that match.

Sixtyfour scores higher on all six measures. On people search, the first result matches on 90.4% of queries against Exa's 72.0%, and precision is 82.8% against 63.3%. On company search, the first result matches on 86.4% of queries against 61.8%. Exa's figures are the results it publishes in its benchmark repository.

Juicebox reports 79% on Exa's people search benchmark. The comparison figures it published beside that number are Exa's precision results, so it compares with Sixtyfour's precision of 82.8%.

Yes. The queries, grading prompts and evaluation harness are open source in Exa's benchmarks repository on GitHub. We graded with GPT-5.4, the default grader in Exa's benchmark scripts until September 2026.

Natural-language people and company searches in the Sixtyfour app, the API and the MCP server. Searches built from filters work as before.

Get started

Create an account and run a natural-language people or company search in the Sixtyfour app.

Sign up
  1. What changed
  2. The results