Open research

Research you can verify, not just read

Harpd publishes open datasets, reproducible benchmark methods and documented methodology. The point is that anyone — a competitor, an auditor, or you — can reproduce our numbers.

Answer

We publish the raw data and the exact methods behind every AI cost and model-quality claim. Where a number is modeled or not yet measured, we say so — we never present estimates as results.

Evidence

What we publish

Everything below is real, machine-readable, and referenced from the pages that make claims.

Open pricing dataset

One JSON/CSV record per model, sourced from official provider pricing pages, each with a sourceUrl and verifiedAt timestamp. The canonical source of truth for every price on Harpd.

Open JSON →

Documented methodology

Four methodologies — model pricing, model replacement, cost per successful task, and reproducible benchmarks — each written down so the measurement is repeatable and auditable.

Read methods →

Reproducible benchmarks

Real-task evaluations with published success definitions, the prices used, and CSV/JSON artifacts. Runner repos are open so the whole run can be replayed.

Benchmark method →

What we don't (yet)

Honesty is the entire point. We are explicit about the gaps.

  • We do not publish measured benchmark results yet — the benchmark is in a PLANNED / MODELED stage (target ship 2026-09-15). Nothing on /benchmarks/ is a result.
  • We do not fabricate model prices. Unknown prices are marked as unavailable, never guessed.
  • We do not present modeled estimates as measured results. Modeled numbers on the site are visually distinct and labeled "modeled".
  • We do not claim a model is "safe to switch" without a real-task quality gate passing on your workload.

Verify it yourself

Pull the dataset, read the method, and re-run the benchmark when it ships. If a Harpd claim doesn't trace back to published data, that's a bug we want to hear about.