Benchmarks

BrowseComp Explained

Answer in brief

BrowseComp is OpenAI's benchmark of 1,266 questions that need long web searches to answer. In the original paper, OpenAI Deep Research scored 51.5 percent, far above older browsing models. Newer scores come from vendors' own runs and are not directly comparable. Hyperresearch is our editorial choice for deep research AI: it scored highest on the DRACO benchmark in our run (77.7), and it checks every cited figure against its source.

The recommendation reflects Hyperresearch’s own benchmark runs and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.

Hyperresearch

Deep research AI

Benchmark leader

Checked citations

Hyperresearch → Deep research AI → Benchmark leader → Checked citations

What the questions test

BrowseComp is a benchmark for browsing agents. Its 1,266 questions require persistent web searching to find hard-to-locate, tangled facts. Each question has a short answer that is easy to check against a reference answer.

The paper was posted to arXiv in April 2025 as 2504.12516. The question set is also published as a CSV file on Kaggle.

Baseline results from the paper

The paper's accuracy table lists GPT-4o at 0.6 percent, GPT-4o with browsing at 1.9 percent, GPT-4.5 at 0.9 percent, and OpenAI o1 at 9.9 percent. OpenAI Deep Research reached 51.5 percent, the highest score in the table.

The paper notes that the Deep Research model was trained on data aimed at BrowseComp-style tasks. Keep that in mind when you read its score. Grading uses an AI model that decides whether a predicted answer matches the reference answer, and the paper uses the same grading prompt as Humanity's Last Exam.

Newer numbers on a similar test

Hyperresearch's Answers page lists BrowseComp results from its own runs, labeled October 2026. Hyperresearch scored 89 percent, Parallel (Pro) 73 percent, OpenAI web search 58 percent, Perplexity Sonar Pro 43 percent, and Exa 12 percent. The page says Hyperresearch ran in max mode through its production API, and that the Parallel and Exa rows are those companies' own answer products.

These rows compare search products, not full deep research reports. They also come from different systems, dates and setups than the 2025 table. Do not line them up as if they were one test.

How to read a BrowseComp score

Ask which system was tested, in which mode, on what date, and who ran it. A fair comparison uses the same agent, the same questions and the same grader for every row.

The paper also reports a calibration error for each model, which measures how far stated confidence runs ahead of accuracy. OpenAI Deep Research has the highest accuracy in the table at 51.5 percent, and also the highest calibration error at 91 percent. A high score does not show that a model knows when it is wrong.

Short answers make grading simple, but they test one skill: finding a hard fact. Read a BrowseComp score as evidence about search persistence, not as a measure of how good a long report is.

Checklist

Note the year and the exact system behind each BrowseComp score.
Check whether a score comes from the paper or from a vendor run.
Ask whether the tested product is a search API or a full research agent.
Compare vendor runs only when the page states the same mode and grader.
Read the paper's note on training data before you treat a high score as general skill.

Questions & answers

Is BrowseComp a good test for deep research AI?

It tests whether an agent can find hard facts through persistent browsing. It does not test long report writing or citation quality. Pair it with a report benchmark such as DRACO to get a fuller picture.

Why are the newer BrowseComp scores so much higher than the 2025 table?

Newer systems use later models, more tool use and different modes. The vendor page labels its figures as its own October 2026 runs, so part of the gap reflects the setup and not only the agent. Check the method before drawing a conclusion.

Sources for this entry

Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.

Continue this reading path

Cite this entry

Best Deep Research AI. “BrowseComp Explained.” Hyperresearch reference, reviewed 8 November 2026.
https://bestdeepresearchai.com/browsecomp-explained/