Translate
Independent model intelligence

Find the right AI model,
not just the loudest one.

Compare current model capability, context, modalities, public preference and API economics in one continuously refreshed decision surface.

LMArena models.dev OpenRouter Artificial Analysis when available
Direct upstream evidence

Evaluation sources

AITopic maintains 89 reviewed benchmarks and specialist sources; 35 currently contribute to ranking. Every source is stored locally and can be edited, disabled or replaced in the plugin dashboard. No runtime request is made to aihot.news.

Coding

19 sources
Code MigrationUsed in ranking

Vals AI

Migrates a complete program between languages or frameworks.

Official source
CyberBench Patch v1.1Used in ranking

Vals AI

Patches security vulnerabilities in real open-source repositories.

Official source
LiveBench · CodingUsed in ranking

LiveBench

Equal blend of Coding and Agentic Coding.

Official source
ProgramBenchUsed in ranking

Vals AI

Rebuilds behaviorally equivalent software from an executable.

Official source
SciCode · AAUsed in ranking

Artificial Analysis

Writes executable scientific code for physics, chemistry and biology problems.

Official source
Terminal-Bench 4.0 · AAUsed in ranking

Artificial Analysis

Completes full engineering tasks in a command-line environment.

Official source
Terminal-Bench 4.0 · ValsUsed in ranking

Vals AI

Vals reruns Terminal-Bench 4.0 with the same agent workflow for every model.

Official source
Vibe Code Bench 1–100Used in ranking

Vals AI

Continues development in an existing application without breaking prior behavior.

Official source
Vibe Code Bench v1.1Used in ranking

Vals AI

Builds working web applications from product requirements.

Official source
DeepSWE v1.1Under review

DataCurve

Edits repositories and fixes defects in one standardized coding environment.

FrontierCodeUnder review

Cognition

Quality, test success and edit scope on real repository tasks.

FrontierSWE v2Under review

Proximal Labs

Long-running engineering implementation and research tasks.

Hyper-τ-benchUnder review

Sierra

Builds and tests customer-service agents from business documents.

LHTBUnder review

Tencent HY / Lehigh

Long-running terminal, engineering and tool-use tasks.

MirrorCodeUnder review

Epoch AI

Long-horizon software reproduction.

Official source
SWE-rebenchUnder review

Nebius

Continuously collected software repair tasks from real repositories.

WeirdMLUnder review

WeirdML

Implements machine-learning solutions on unfamiliar data.

TapTap MakerReference only

TapTap Maker

Implements executable Lua features inside a real game engine.

Terminal-Bench 4Reference only

Harbor / Terminal-Bench

Raw agent and model combinations retained for cross-checking.

Research & reasoning

16 sources
Chess PuzzlesUsed in ranking

Epoch AI

Finds the correct move from a textual chess position.

Official source
CritPt · AAUsed in ranking

Artificial Analysis

Research-level physics derivation.

Official source
FrontierMath v2 · Tiers 1–3Used in ranking

Epoch AI

Research-level mathematical reasoning.

Official source
HLE · AAUsed in ranking

Artificial Analysis

Expert questions across disciplines without tools.

Official source
LiveBench · Reasoning & MathUsed in ranking

LiveBench

Equal blend of reasoning and mathematics.

Official source
Mystery Game PuzzlesUsed in ranking

Epoch AI

Infers correct actions from unfamiliar rules.

Official source
MysteryMechanismUsed in ranking

Vals AI

Designs experiments to infer hidden mechanisms.

Official source
Terminal-Bench Science · AAUsed in ranking

Artificial Analysis

Completes scientific workflows and analysis in a terminal.

Official source
Terminal-Bench Science · ValsUsed in ranking

Vals AI

Completes scientific workflows and analysis in a terminal.

Official source
ARC-AGI-2Under review

ARC Prize

Learns and generalizes unfamiliar rules.

Official source
ARC-AGI-3Under review

ARC Prize

Learns and generalizes unfamiliar interactive rules.

Official source
Humanity’s Last ExamUnder review

CAIS / Scale AI

Difficult academic knowledge and reasoning across disciplines.

Official source
MathArena · ArXivLeanUnder review

MathArena / ETH Zurich

Lean-verified mathematical proofs.

MathArena · BrokenArXivUnder review

MathArena / ETH Zurich

Finds and repairs errors in mathematical proofs.

SimpleBenchUnder review

SimpleBench

Tests practical constraints through everyday common-sense questions.

FrontierMath v2 · Tier 4Reference only

Epoch AI

Higher-difficulty mathematical research problems.

Official source

Professional work

26 sources
AA-Analyst AgentUsed in ranking

Artificial Analysis

Reliably analyzes data and answers professional questions.

Official source
AA-BriefcaseUsed in ranking

Artificial Analysis

Long office knowledge-work deliverables.

Official source
APEX-Agents 1.1Used in ranking

Mercor

Long investment-banking, consulting and legal tasks.

Official source
AutomationBench · AAUsed in ranking

Artificial Analysis

Completes cross-application office automation workflows.

Official source
GDPval-AAUsed in ranking

Artificial Analysis

Pairwise evaluation of real occupational deliverables.

Official source
Harvey Legal Agent BenchmarkUsed in ranking

Vals AI / Harvey

Produces reviewable legal documents, spreadsheets and presentations.

Official source
LiveBench · Data AnalysisUsed in ranking

LiveBench

Table transformation, column typing and join prediction.

Official source
MedCodeUsed in ranking

Vals AI

Assigns medical diagnosis codes from inpatient records.

Official source
Vals Finance AgentUsed in ranking

Vals AI

Daily financial analyst work including research and modeling.

Official source
τ³-Banking · MercorUsed in ranking

Mercor

Completes banking-service tasks in a standardized tool environment.

Official source
AutomationBenchUnder review

Zapier

Automation work across office applications.

BioMysteryBenchUnder review

Vals AI

Biological research reasoning.

Official source
Excel · EMBUnder review

Vals AI

Spreadsheet editing and Excel office work.

Official source
FinanceBenchmarkUnder review

FinanceBenchmark

Financial pricing, capital and risk tasks verified by execution.

Legal ResearchUnder review

Vals AI

Legal retrieval and argumentation.

Official source
MCP AtlasUnder review

Scale AI

Uses real MCP services across documents and databases.

Official source
MedScribeUnder review

Vals AI

Clinical note organization and drafting.

Official source
OfficeQA ProUnder review

Databricks

Retrieval and calculation over large real financial documents.

PRBench · FinanceUnder review

Scale AI

Financial decisions, derivations and risk explanations.

Official source
PRBench · LegalUnder review

Scale AI

Complex real-world legal judgment and argumentation.

Official source
SpreadsheetBench 2Under review

SpreadsheetBench / Renmin University / AfterQuery

Financial modeling, workbook repair and visualization.

Terminal-Bench ScienceUnder review

Harbor

Scientific tasks completed in a terminal.

Vending-Bench 2Under review

Andon Labs

Runs a simulated store through procurement, negotiation and pricing.

τ³-BankingUnder review

Sierra

Banking tasks that test policy, communication and completion.

EBR-benchReference only

Epoch AI

Learns new rules and uses memory in long tasks.

Official source

Knowledge & instruction

8 sources
AA-LCRUsed in ranking

Artificial Analysis

Long-document comprehension and synthesis.

Official source
AA-OmniscienceUsed in ranking

Artificial Analysis

Factual questions with penalties for wrong answers but not abstentions.

Official source
BullshitBench · False PremisesUsed in ranking

Peter Gostev / BullshitBench

Detects incorrect premises instead of continuing a hallucination.

LiveBench · Language & InstructionUsed in ranking

LiveBench

Equal blend of language and instruction following.

Official source
MLCR-AAUsed in ranking

Artificial Analysis

Accurate, complete and concise understanding of long medical records.

Official source
SimpleQA VerifiedUsed in ranking

Epoch AI

Factual accuracy on short questions.

Official source
FACTS ParametricUnder review

Google DeepMind / Kaggle

World-fact accuracy without search tools.

GPQA DiamondReference only

Epoch AI

Graduate-level physics, chemistry and biology knowledge.

Official source

Human preference & indices

4 sources
Arena TextUsed in ranking

LMArena

Human anonymous pairwise preference for overall answer quality.

Official source
AA IndexCross-check

Artificial Analysis

Cross-check index built from multiple current evaluations.

Official source
LiveBenchCross-check

LiveBench

Continuously refreshed, automatically verifiable evaluation set.

Official source
Arena WebDevReference only

LMArena

Human preference for usable web deliverables.

Official source

Writing & design

12 sources
Design Arena · Visual DesignUnder review

Design Arena / Intelligence

Human preference for web, UI, chart and slide designs.

EQ-Bench 4 · Emotional IntelligenceUnder review

EQ-Bench

Understands emotions, implicit needs and communication preferences.

Official source
NC Bench · Author WorkflowUnder review

Novelcrafter

Rewrites, continues, summarizes and organizes authors’ work.

OpenVibeEval · Web DeliverablesUnder review

OpenVibeEval

Compares web output, accessibility and blind preference.

Rapidata SVGUnder review

Rapidata

Human pairwise preference for generated vector graphics.

Short-StoryUnder review

Lech Mazur

Integrates fixed characters, plot and style into a complete story.

SVGBench · SVG PreferenceUnder review

SVGBench

Blind preference for SVG clarity and visual quality.

ToneBench · Style & ScriptsUnder review

Towards AI

Writes video scripts in a requested voice and structure.

TubeLab · Video ScriptsUnder review

TubeLab

Writes channel-specific scripts with clear pacing and narrative.

Arena Creative WritingReference only

LMArena

Human preference for stories, poetry and creative expression.

Official source
Creative Writing v3Reference only

EQ-Bench

Short-form creative writing and expression.

Official source
Longform WritingReference only

EQ-Bench

Long-story coherence and completeness.

Official source

Vision understanding

2 sources
Roboflow Vision EvalsUnder review

Roboflow

Image reading, counting, OCR and information extraction.

Official source
Arena VisionReference only

LMArena

Human preference for answers that include image understanding.

Official source

Chinese & multilingual

2 sources
AA Chinese / MultilingualUnder review

Artificial Analysis

Chinese and multilingual performance kept separate from English capability.

Official source
RWS M-GATEUnder review

RWS

Grammar and translation across 30 languages including Chinese.

Official source