ResearchLLM Verifications Leaderboard

Data as of · 78 models, 80 questions · reviewed by Suhas Bhairav · methodology

Data

Download the benchmark data

The numbers behind every page, one row per model, in CSV or JSON. The files are rebuilt from the latest run, so the same URL always holds the current data.

Maintained by Suhas Bhairav · Data as of

Files

The JSON file also carries the suite name and hash, the source datasets with their licences, and the number of speed runs per model.

Columns

Empty cells mean the measure was not taken or not published for that model (for example, no score when too many requests failed on the host's side). Every definition is on the methodology page.

model_id, name, creator, url, open_weights, reasoning, score_pct, correct, graded, failed_requests, cost_per_question_usd, cost_per_correct_usd, output_tokens_per_s, time_to_first_token_s, end_to_end_s, output_tokens_per_question, context_window, list_input_usd_per_m, list_output_usd_per_m, quality_tested_at, speed_tested_at, supergpqa_correct, mmlu-pro_correct, subject_science_correct, subject_engineering_correct, subject_medicine_correct, subject_economics_correct, subject_law_correct, subject_humanities_social_science_correct

Using the data

You may reuse the numbers with a link back to the page they come from. The questions themselves belong to their datasets and keep their own licences: SuperGPQA (ODC-BY 1.0), MMLU-Pro (MIT). Scores come from 80 questions, so a gap of a few points between two models is usually within chance.

Suggested citation

Suhas Bhairav. LLM Verifications Leaderboard AI model benchmark results, data as of 2026-10-27. https://suhasbhairav.com/research/llm-leaderboard/downloads