AI Benchmarks & Leaderboards

Public product-experience rating normalized from 4.0/5 to 80.0/100.

Publicly reported reasoning benchmark result.

Low-weight product-experience proxy normalized from a public 5-star parent/platform rating.

MMMU — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Public product-experience rating normalized from 1.7/5 to 34.0/100.

Public product-experience rating normalized from 5.0/5 to 100.0/100.

Structured performance benchmark tracked by AI Orbit.

Public product-experience rating normalized from 4.1/5 to 82.0/100.

AIME 2025 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Humanity's Last Exam (No Tools) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Aider Polyglot (Diff) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

ERQA — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMMU-Pro (Average Standard + Vision) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

VideoMMMU (Max 256 Frames) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

CharXiv Reasoning (Python Enabled) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Public product-experience rating normalized from 4.7/5 to 94.0/100.

SWE-bench Verified (OpenAI 477 Tasks) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

HMMT 2025 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Publicly reported mathematical reasoning benchmark result.

Publicly reported professional agents benchmark result.

Publicly reported software engineering benchmark result.

Independent provider-voice speech arena; Elo reflects pairwise listener preference for generated speech quality.

Publicly reported abstract reasoning benchmark result.

Publicly reported computer-aided design benchmark result.

Publicly reported cybersecurity benchmark result.

Publicly reported cybersecurity benchmark result.

FrontierMath (Python tool only) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Controlled code-review benchmark across 50 hard-to-find bugs; F1 balances precision and recall.

Living open-source code-review benchmark based on real-world pull-request behavior; F1 balances precision and recall.

Blind pairwise human-preference Image Arena quality rating. Bradley-Terry maximum-likelihood ratings are rescaled to an Elo-like scale.

DeepResearch Bench RACE overall report-quality score across 100 PhD-level tasks in 22 fields.

Agents' Last Exam — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Toolathlon — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-Bench Pro — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Terminal-Bench 2.1 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

OSWorld 2.0 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Capture-the-Flag Challenges — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

ExploitGym — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

HealthBench Professional — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

GraphWalks BFS 1M F1 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

GraphWalks BFS 256K F1 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

OpenAI MRCR v2 8-Needle 256K-512K — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

OpenAI MRCR v2 8-Needle 512K-1M — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

FrontierMath Tier 1-3 v2 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMMU Pro (No Tools) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMMU Pro (With Tools) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Big Finance Bench — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Internal Research Debugging Evaluation — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

KernelGen 1P — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

PostTrainBench Lite — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

GeneBench Pro — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

LifeSciBench — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

BrowseComp — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Independent full-stack app-builder composite measured with the same nine-prompt specification across five runs (45 recorded executions per product), scored 0–100.

Aider Polyglot — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

LiveCodeBench — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SimpleQA — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMMLU — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMLU — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

KLAS Research 100-point overall performance score based on healthcare-provider feedback.

SWE-bench Verified (Anthropic Full 500 High Compute) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-bench Verified (Anthropic Full 500) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-bench Verified (Google Multiple Attempts) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-bench Verified (Google Single Attempt) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Publicly reported coding & terminal agents benchmark result.

FACTS Grounding — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

LOFT Hard <=128K — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

LOFT Hard 1M — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MRCR v2 8-Needle <=128K — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MRCR v2 8-Needle 1M — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Global MMLU Lite — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMLU-Pro is a robust multi-task language-understanding benchmark with a stronger emphasis on challenging reasoning questions.

Publicly reported scientific agents benchmark result.

ActivityNet-QA — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

EgoTempo — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Perception Test — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

QVHighlights — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

VideoMMMU — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

PINT evaluates prompt-injection detection systems on a 4,314-input multilingual dataset; reported score is higher-is-better, with benchmark configurations documented publicly.

Blind pairwise human-preference Music Arena rating; users compare tracks from the same prompt and modality, with ratings calculated on an Elo-like scale.

Blind pairwise human-preference Music Arena rating; users compare tracks from the same prompt and modality, with ratings calculated on an Elo-like scale.

Human-validated 500-instance SWE-bench subset used to evaluate coding agents and systems.

Terminal-Bench 2.0 leaderboard accuracy for terminal-based agent tasks. Results are listed with agent, model, date and organization.

Independent live speech-to-text benchmark on pinned, versioned datasets under production-realistic conditions. Word Error Rate (WER) is lower-is-better.

Independent pairwise preference arena for AI video editing with audio; Elo is a relative quality rating.

Independent pairwise preference arena for image-to-video generation with audio; Elo is a relative quality rating.

Independent pairwise preference arena for text-to-video generation with audio; Elo is a relative quality rating.

ARC-AGI-2 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Tau2 Average — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

C-Eval — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

C-SimpleQA — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

CLUEWSC — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Codeforces Percentile — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

LiveCodeBench v6 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-bench Verified (Anthropic 489 High Compute) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-bench Verified (Anthropic 489 Internal Subset) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

SWE-bench Verified (DeepSeek Resolved) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Terminal-bench — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Publicly reported cybersecurity benchmark result.

Publicly reported cybersecurity / reliability benchmark result.

Publicly reported cybersecurity patching benchmark result.

Publicly reported finance agents benchmark result.

IF-Eval (Prompt Strict) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

Publicly reported knowledge & reasoning benchmark result.

Publicly reported knowledge & reasoning benchmark result.

Publicly reported legal agents benchmark result.

MRCR v2 8-Needle 128K Average — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

AIME 2024 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

AIME 2026 (No Tools) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

CNMO 2024 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MATH-500 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMMU Pro — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

AlpacaEval2.0 — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

ArenaHard — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

BigBench Extra Hard — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

DROP (3-shot F1) — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

FRAMES — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MMLU-Redux — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

MATH-Vision — official/provider-reported evaluation result. Methodology conditions are preserved in the result notes.

IIHS real-world safety comparison of Waymo driverless operation with human drivers in matched locations/periods; score is percent lower police-reportable crash involvement per mile.

Tesla's rolling Vehicle Safety Report comparison for major collisions with FSD (Supervised).

Clinical evaluation of Viz.ai-LVO against board-certified neuroradiologist CTA interpretation in a large integrated stroke network; score used here is overall accuracy.

Independent Nature Medicine evaluation on 500 MedQA questions comparing specialized clinical AI tools and frontier models; score is answer accuracy.

Independent pairwise preference arena restricted to open-weight text-to-image models; Elo is a relative quality rating.

Vals Legal AI Report (VLAIR) independent evaluation of legal AI applications on realistic legal work; score used here is Document Q&A accuracy.

Real-world production-time reduction calculated from an official customer example: two weeks to five days.

Public product-experience rating normalized from a 5-star scale to 0–100 for AI Orbit display/import.

Public product-experience rating normalized from a 5-star scale to 0–100 for AI Orbit display/import.

Product-experience rating reported on the official Copy.ai reviews page and attributed there to G2.

Low-weight product-experience proxy normalized from a 5-star parent/platform rating.

Sensitivity/recall for abstract screening on the published Cochrane-derived evaluation set.

Percent correct for Methods, Participants and Interventions extraction tasks on the published evaluation.

Paper-level recall at the full-text screening stage on the published Cochrane-derived evaluation.

Recall of final-included studies retrieved using only the systematic-review title as the query.

You.com public ARI evaluation on the independently developed FRAMES benchmark; score uses the original FRAMES evaluation method.

Figure's public controlled evaluation of Helix S1 logistics manipulation after scaling demonstration data; score used here is barcode-orientation success rate.

Official NEO tendon-drive backdrivability specification.

Official Boston Dynamics Spot maximum locomotion speed specification.

Observed unloading throughput reported in an official Boston Dynamics customer case study.

Accuracy on Exa's open WebCode Contents track measuring extraction fidelity against golden markdown.

YouGov net Satisfaction score among U.S. current and former users of the AI brand.