Gender Prediction API Benchmark 2026

How NameGender compares

Every provider below answered the same names, with the same country hint, on the same day. A name that any provider failed to return at all — a network or quota error, not an unknown answer — was dropped from every provider's table, so all rows share one denominator. Unknown answers are each provider's own decision; no threshold is imposed on anyone.

Confident errors are wrong answers returned with a probability of 0.90 or higher — the mistakes a probability threshold cannot filter out for you. Where two confidence intervals overlap, the ranking between those two rows is not a result.

By writing system

An overall average hides where providers actually differ. The gaps between them are widest outside the Latin alphabet.

Measured on 991 names shared by every provider. Answers were collected on different days: NameGender October 6, 2026; genderize.io September 18, 2026; NamSor September 18, 2026. Competitor answers are aggregated only: this page publishes no per-name result from another provider. Our own wrong answers are listed in the report file so the score can be checked against the fixture. Prices, limits and behaviour change; re-run the measurement before quoting it.

Which provider fits which job

Each writing system has 40 to 300 names in this test, so a gap of a few points is within the noise. Gaps of ten points or more are not.

Where NameGender gets it wrong

Labels come from the test set, which assigns one gender per name. Some of these names are given to both genders in practice; the table lists them as errors anyway.

Run the test yourself

The script below sends the published test names to the NameGender API, country by country, and prints coverage, answered accuracy and end-to-end accuracy. Point it at another provider's endpoint to compare on your own terms. It needs Python 3 and the requests package, and uses one credit per name.

import csv, io, os, requests
from collections import defaultdict

names_csv = requests.get("https://namegender.com/datasets/gender-api-benchmark-2026-names.csv").content.decode("utf-8-sig")
rows = list(csv.DictReader(io.StringIO(names_csv)))

by_country = defaultdict(list)
for row in rows:
    by_country[row["country"]].append(row)

answered = correct = 0
for country, group in by_country.items():
    for start in range(0, len(group), 100):
        batch = group[start:start + 100]
        reply = requests.post(
            "https://namegender.com/api/v1/gender/bulk",
            headers={"Authorization": f"Bearer {os.environ['NAMEGENDER_API_KEY']}"},
            json={"names": [r["name"] for r in batch], "country": country},
            timeout=60,
        ).json()
        for row, result in zip(batch, reply["results"]):
            if result["gender"]:
                answered += 1
                correct += result["gender"] == row["expected_gender"]

total = len(rows)
print(f"coverage {answered / total:.1%}")
print(f"answered accuracy {correct / answered:.1%}")
print(f"end-to-end {correct / total:.1%}")

A free account includes 100 credits a day. Researchers who want to rerun or extend the comparison can apply for research credits.

Turkish names: a separate 1,317-name test

• Turkish test names and results (CSV) — every name with its label, the number of people behind it, whether it uses Turkish letters and NameGender's answer (empty means unknown). CC BY 4.0; the Wikidata records are CC0.

Measured October 22, 2026 on the live engine · names built October 22, 2026 · fuzzy matching enabled

Results by country

The aggregate is not portable to every market. Israel and Thailand, for example, expose coverage gaps that the overall figure cannot show.

Results by writing system

Data and citation

Two files let you check the vendor comparison above or rerun it against any service, including ours. The labelled names come from the MIT-licensed Faker library; the files are free to use under CC BY 4.0.

• Test names (CSV) — every name in the comparison with its country, expected gender and writing system.
• NameGender misses (CSV) — every name NameGender got wrong or left unanswered, with its answer and probability. An empty answer means unknown.

Genderize and Namsor answers were recorded as totals per group, not name by name, so only their aggregated results are published.

How to cite

NameGender (2026). Gender Prediction API Benchmark 2026. https://namegender.com/research/gender-api-benchmark-2026

@misc{namegender_benchmark_2026,
  author       = {NameGender},
  title        = {Gender Prediction API Benchmark 2026},
  year         = {2026},
  howpublished = {\url{https://namegender.com/research/gender-api-benchmark-2026}}
}

How to interpret this report