Registry / benchmarks
37 benchmarks · 18 with scores · 2,694 score observations · 169 source documents
tracking since 2026-10-02 · last ingest 2026-10-05 · live dataset
Every benchmark's scores are tracked per source and never merged into a cross-source ranking — different sources measure under different conditions, and each row cites the document that reported it.
Scores are re-collected from every source on each ingest, and first-seen dates are preserved — each benchmark page doubles as a long-term tracking record. Score observations and source documents located via Benchmark Radar (benchmark-radar.org; code MIT, content CC BY-NC-SA 4.0, non-commercial use with attribution). Every row cites its primary source document; all descriptions are AgentX's own. Sources are kept as separate partitions and never merged into cross-source rankings.
37 of 37 benchmarks
| Benchmark | Scores reported | Best score | Released | Repo |
|---|---|---|---|---|
| GPQA Knowledge & QA34 docs | 859 801 models | 94.9% Artificial Analysis | Nov 2023 | idavidrein/gpqa |
| Humanity's Last Exam Knowledge & QA24 docs | 708 677 models | 55.5% Artificial Analysis | Jan 2025 | centerforaisafety/hle |
| SciCode Research & Discovery4 docs | 600 594 models | 60.2% Artificial Analysis | Jul 2024 | scicode-bench/SciCode |
| SWE-bench Coding & SWE20 docs | 129 117 models | 95% LLM Stats | Aug 2024 | SWE-bench/SWE-bench |
| Terminal-Bench 2.0 Coding & SWE23 docs | 99 87 models | 88.8% LLM Stats | May 2025 | harbor-framework/terminal-bench-2 |
| Tau2-Bench Tool Use9 docs | 89 41 models | 99.3% LLM Stats | Jun 2025 | sierra-research/tau2-bench |
| OSWorld Computer Use8 docs | 48 47 models | 86.1% LLM Stats | Apr 2024 | xlang-ai/OSWorld |
| Toolathlon Tool Use11 docs | 47 42 models | 75.6% LLM Stats | Nov 2025 | hkust-nlp/Toolathlon |
| MCP Atlas Tool Use9 docs | 41 36 models | 88.1% LLM Stats | Nov 2025 | scaleapi/mcp-atlas |
| SuperGPQA Knowledge & QA5 docs | 38 37 models | 73.6% LLM Stats | Feb 2025 | SuperGPQA/SuperGPQA |
| Claw-Eval Computer Use1 doc | 14 14 models | 80.9% LLM Stats | – | claw-eval/claw-eval |
| ClawBench Computer Use1 doc | 8 8 models | 33.3% clawbench paper 153 tasks · system cards | Apr 2026 | TIGER-AI-Lab/ClawBench |
| PinchBench General Agents1 doc | 6 6 models | 90% LLM Stats | – | pinchbench/skill |
| PaperBench Research & Discovery2 docs | 3 3 models | 93% LLM Stats | Apr 2025 | openai/frontier-evals |
| MLE-bench Research & Discovery3 docs | 2 2 models | 2 reports | Oct 2024 | openai/mle-bench |
| BixBench Bio & Medicine1 doc | 1 1 models | 1 report | Feb 2025 | Future-House/BixBench |
| Harbor Coding & SWEHarness1 doc | 1 1 models | 1 report | Sep 2026 | harbor-framework/harbor |
| ResearchClawBench Research & Discovery2 docs | 1 1 models | 1 report | – | InternScience/ResearchClawBench |
| ActiveSciBench Research & Discovery1 doc | – 0 models | – | May 2026 | scientific-discovery/LLM-AutoSciLab |
| AgentBench General Agentsnot yet tracked | – 0 models | – | Aug 2023 | THUDM/AgentBench |
| AutoMedBench Bio & Medicine1 doc | – 0 models | – | Jun 2026 | AutoMedBench/AutoMedBench |
| BioAgent Bench Bio & Medicine1 doc | – 0 models | – | Jan 2026 | bioagent-bench/bioagent-bench |
| BioDSA-1K Bio & Medicinenot yet tracked | – 0 models | – | May 2025 | huggingface.co/zifeng-ai/BioDSA-1K |
| ClawMark General Agentsnot yet tracked | – 0 models | – | – | evolvent-ai/ClawMark |
| DrugDiscoveryBench Chemistry & Drugsnot yet tracked | – 0 models | – | Sep 2026 | scaleapi/DrugDiscoveryBench |
| GENEB Bio & Medicine1 doc | – 0 models | – | Jun 2026 | darlednik/GENEB |
| HealthAgentBench Bio & Medicine1 doc | – 0 models | – | Jun 2026 | microsoft/HealthAgentBench |
| HeurekaBench Bio & Medicinenot yet tracked | – 0 models | – | Jan 2026 | mlbio-epfl/HeurekaBench |
| Holistic Agent Leaderboard General Agentsnot yet tracked | – 0 models | – | Oct 2025 | princeton-pli/hal-harness |
| HybridDeepResearch General Agents1 doc | – 0 models | – | Sep 2026 | Snowflake-AI-Research/HybridDeepResearch |
| Longevity Claw Research & Discoverynot yet tracked | – 0 models | – | Sep 2026 | Insilico-org/longeclaw |
| MedAgents-Bench Bio & Medicine1 doc | – 0 models | – | Mar 2025 | gersteinlab/MedicalAgentsBench |
| PatientAgentBench Bio & Medicine1 doc | – 0 models | – | Jul 2026 | amazon-science/PatientAgentBench |
| RE-Bench Research & Discovery1 doc | – 0 models | – | Nov 2024 | METR/RE-Bench |
| RareArena Bio & Medicinenot yet tracked | – 0 models | – | Feb 2026 | zhao-zy15/RareArena |
| ScienceAgentBench Coding & SWEnot yet tracked | – 0 models | – | Oct 2024 | OSU-NLP-Group/ScienceAgentBench |
| deep_research_bench Research & Discovery1 doc | – 0 models | – | Jun 2025 | Ayanami0730/deep_research_bench |
“Scores reported” counts every recorded value with its source document — the tracking history, not a leaderboard. “Best score” is the top value in a benchmark’s largest source partition; sources are never merged. The full history lives on each benchmark page and in the public API.