Benchmarks

Investment Banking

Financial analysis over filings and credit agreements — valuation math, document QA, and the quantitative reasoning behind deals.

DeepSeek V4-Flash87.2

Corporate Law

Legal reasoning — issue spotting, rule application, and contract analysis — plus the broad knowledge a generalist counsel needs.

GPT-5.486.0

Grok 4.384.5

Claude Sonnet 4.682.1

Nova 2 Pro81.6

Claude Haiku 4.581.2

Medicine

Clinical knowledge and diagnostic reasoning, including the medical coding accuracy and science depth that real practice demands.

DeepSeek V4-Flash87.2

Scientific Research

Frontier problem solving — graduate-level science, the hardest multi-domain exams, and broad expert knowledge.

Management Consulting

Broad analytical reasoning across business domains — structured problem solving over wide-ranging knowledge.

Accounting & Audit

Numerically exact work over tax and financial documents — reconciliation, controls, and the arithmetic discipline audits demand.

Nemotron 3 Ultra86.8

DeepSeek V4-Flash86.2

Qwen3.6-35B-A3B85.2

Nova 2 Pro81.6

DeepSeek V4-Pro73.7

Language Models

Reasoning23 results

GPQA Diamond

Graduate-level, Google-proof science questions (physics, chemistry, biology) written by domain experts to resist web lookup.

Frontier22 results

Humanity's Last Exam

A broad, extremely difficult exam across math, humanities, and science designed to remain unsaturated by frontier models. Reported here without external tools.

Agentic Coding20 results

SWE-bench Verified

A human-validated subset of real GitHub issues; the model must produce a patch that passes the repository's tests. Figures are vendor-reported unless noted.

Coding13 results

LiveCodeBench

Contamination-resistant competitive-programming problems collected over time to avoid training-set overlap.

Knowledge10 results

MMLU-Pro

A harder, cleaned-up successor to MMLU spanning 57+ subjects with 10-way multiple choice and reasoning-heavy items.

Human Preference25 results

LMArena Elo

Crowd-sourced pairwise preference rating from blind head-to-head chats (LMArena, formerly LMSYS). Unbounded; ~1000 is the historical anchor.

elo · higher is better

Composite31 results

AA Intelligence Index

Artificial Analysis Intelligence Index (v4.0) — an independent composite across ~10 evaluations (incl. GPQA Diamond, HLE, Terminal-Bench, SciCode, GDPval, τ²-Bench). The de-facto cross-model standard; higher is better. Shown for reference and normalized relative to this set.

pts · higher is better

Finance26 results

Corporate Finance

Vals AI CorpFin v2 — expert-built questions over long-context corporate credit agreements; an independent, in-house-run finance benchmark.

Law24 results

LegalBench

Legal-reasoning task suite (originated by Stanford CodeX), run independently by Vals AI and reported as overall accuracy across tasks.

Tax & Accounting25 results

TaxEval

Vals AI TaxEval v2 — 1,500+ expert-written tax questions, scored on overall accuracy. Independent, in-house-run.

Medicine21 results

Medical Coding

Vals AI MedCode — accuracy of ICD-10-CM diagnosis coding for the medical billing process. Independent, expert-built dataset.

Physical Prediction0 results

World Models

emerging

Physion++

Predict whether physical scenarios resolve as expected (will objects collide, fall, or stay stable?). Probes intuitive physics.

Action Anticipation0 results

EK-100 Anticipation

EPIC-Kitchens-100 long-term action anticipation — forecast the next actions in egocentric video. Reported as mean top-5 recall.

recall · higher is better

Generative Coherence0 results

Video Prediction (FVD)

Fréchet Video Distance between predicted and ground-truth future frames. Measures rollout realism — lower is better.

FVD · lower is better

Planning0 results

Embodied Planning

Success rate of model-based planning / imagined rollouts on embodied control and navigation tasks.