Data as of · 78 models, 80 questions · reviewed by LLM Verifications team · methodology

Editorial standards

Editorial standards

The rules the LLM Verifications team follows for every number, ranking and page on LLM Verifications.

Reviewed by LLM Verifications team · Last updated

Independence

  • Rankings and verdicts come only from our own measurements. No model maker or API host can pay for placement, ranking, inclusion or a review.
  • There are no sponsored results and no affiliate links in rankings.
  • We don't accept model submissions. The team decides which models to test.
  • No lab or host sees results before they are published.

Sourcing

What we measure ourselves

Verification scores, output speed, time to first token, end-to-end time, tokens per question and billed cost are all measured by our harness, from requests we sent and recorded. None of them is copied from a model card, vendor blog post or another leaderboard.

What we take from others

  • Questions come from openly licensed datasets, credited with links to the dataset and its paper on the benchmarks page and in the methodology.
  • Specifications (context window, release date, list prices, the hosts that serve a model, open-weights status, knowledge cutoff) come from OpenRouter's public model catalogue and are labelled as such.

Review before publishing

Every benchmark run is reviewed by the LLM Verifications team before it replaces the published data. The review checks that:

  • each model answered the full question suite, and that host-side failures stayed under the 20% limit above which a model gets no score;
  • speed results are plausible, and that models that deliver their answer in a single burst after hidden thinking have their speed withheld;
  • billed costs agree with the listed price of the host that served each request, or that any gap has an explanation;
  • the suite and harness hashes match what the run was meant to use.

Results are never edited by hand. If a run has a problem, we fix the cause and re-run the affected models.

Corrections

If you think a number is wrong, contact us at support@globesword.com with the page and the model. The team checks the recorded request and response behind it and replies. When we find an error, we correct it by re-running the affected tests, and note material changes, such as a changed method or a removed result, in the changelog.

Automation and AI

The data pages on this site (leaderboards, model, provider, host, subject and comparison pages, including their summaries and quick answers) are generated by code directly from the measured data. Each sentence is a template filled with measured numbers, so it changes when the data changes, and no language model writes or edits the published figures. Grading is automatic: answers are matched against the dataset's answer key, without a human or AI judge. The LLM Verifications team writes and reviews the methodology and the other editorial pages, and reviews each run before it is published.

Conflicts of interest

The site has no commercial relationship with the labs or hosts it ranks: no sponsorship, no paid placement and no referral fees. If that ever changes, we will say so on this page and on the pages affected before it takes effect. We disclose that every model is reached through OpenRouter, which means OpenRouter's routing and pricing are part of what we measure.

See also about us and the methodology.