Every public benchmark, one place to run it

Pick benchmarks, models and a config. Benchmark Router runs the official benchmark code and returns metrics and full logs, on any model or your own harness.

Benchmarks you can run on Benchmark Router include Humanity's Last Exam, Harvey LAB, Terminal-Bench 4.0, HealthBench, BankerToolBench, SWE-bench Multimodal, CyBench, ParseBench 2, EnterpriseOps-Gym, RedlineBench, PlanBench-V, DeepSWE, AutomationBench, CVE-Bench, HLE: Diamond, ProgramBench.

How it works

Pick, run, hill climb

try benchmark router ↗

Benchmark Router sits between your model or harness and the official benchmark code, and sends every result to one place.

General

KnowledgeScienceVision

Coding & agentic

CodingAgentic

Industry

LegalFinanceHealth

import newmeasure as nm
router = nm.BenchmarkRouter()run = router.runs.create(
    benchmarks=["hle", "cybench"],
    models=["claude-opus-5.5"],
    config={"trials": 3},
)run.wait()  # official code, meteredprint(run.summary())
# metrics · cost per task
# logs/ · one trace per task

Trust

A score you can launch with

New Measure independently runs and audits the official benchmarks on your product, so buyers don't have to take your word for the score. Publish your position with a third party behind it.

  • Official code

    Official benchmark code and grading, run on your harness, not just the base model. Same scale as published scores.

  • Full logs

    The trace of every task: what the model saw, what it answered, how it was graded. Anyone can check the score.

  • Verified and published

    Audited results, hosted on your own domain with a Verified by New Measure mark. Ready for launch day.

Tell us what to benchmark

Benchmark Router is live. Run public benchmarks yourself, or tell us what's missing and we'll add it or run it for you.

  • Public benchmarks, ready to run
  • New benchmarks added on request
  • Verified results you can publish