Benchmark By Score Rank
Overall 1
AA Intelligence Index4.3.2 Artificial Analysis combines tests of knowledge, reasoning, coding and tool use into one score AA 49.6 #10 of 337
Knowledge 7 Leads on 1
AA OmniscienceAccuracy Measures correct answers on the AA Omniscience factual knowledge test AA 65.4% #3 of 321+1 versions & slices Hallucination Rate Measures how often attempted answers are incorrect on the AA Omniscience factual knowledge test AA 63.6% #116 of 321
AA Omniscience Index Scores factual knowledge while penalizing incorrect answers to encourage admitting uncertainty AA 43.3 #4 of 321
IFBench (instructions) Checks whether the model follows precise formatting and content instructions AA 63.5% #75 of 248
EQ-Bench3 Judge-rated responses to social and emotionally nuanced scenarios, scored through the source pairwise ranking EQ-Bench 2,050 #1 of 50
SimpleQA Verifiedv1 Accuracy on short factual questions with verified answers, answered without web access 70.7% #8 of 83
Mystery Game Puzzles Models choose the best move in puzzle positions from an undisclosed game Epoch 52% #11 of 64
Reasoning 5
ARC-AGI-1public evaluation Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 97.9% #12 of 61+1 versions & slices semi-private Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 98.5% #1 of 64
ARC-AGI-2public evaluation Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 93.3% #8 of 63+1 versions & slices semi-private Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 88.3% #10 of 66
Chess Puzzles Models choose the best next move from a novel chess position Epoch 47% #19 of 68
Math 11
AA Analyst Agent Data-analyst tasks: the model explores data with tools and reports findings AA 48.8% #10 of 43
FrontierMath Erdos Models solve expert-written advanced mathematics problems on the named tier and edition Epoch 0% #6 of 13
FrontierMath Tier 4 (v2) Models solve expert-written advanced mathematics problems on the named tier and edition Epoch 90.2% #4 of 54
FrontierMath Tiers 1–3 (v2) Models solve expert-written advanced mathematics problems on the named tier and edition Epoch 87% #9 of 69
MathArenaArXivMath · April 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 72.5% #3 of 17+6 versions & slices ArXivMath · June 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 85.4% #6 of 24ArXivMath · May 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 87.5% #4 of 19BrokenArXiv · April 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 54.1% #4 of 15BrokenArXiv · June 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 47.8% #13 of 24BrokenArXiv · May 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 44.5% #7 of 18ArXivLean · June 2026 Answer accuracy on the named mathematics competition under the recorded effort MathArena 15.2% #6 of 9
Science 3
Humanity's Last Exam Very hard expert questions across many fields, written to stump today’s models AA 55.5% #4 of 337
CritPt (physics) Research-level physics problems that need many careful reasoning steps AA 28.6% #13 of 324
SciCode Writing code for real scientific computing problems (physics, biology, math) AA 61% #5 of 140
Coding 7
Terminal-Bench4.0 Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite 42.4% #11 of 132+1 versions & slices 2.1 Older, easier version of Terminal-Bench; Still widely reported AA 84.6% #12 of 156
DeepSWE1.1 Agents change repository code to resolve software engineering tasks under the stated DeepSWE edition 69.9% #9 of 44
WebDev Arena Human pairwise preferences in the stated webdev vote pool Arena 1,624 #20 of 97+1 versions & slices full cohort Human pairwise preferences in the stated webdev vote pool Arena 1,624 #20 of 97
MirrorCode Agents reimplement software projects by observing their behavior Epoch 63.9% #3 of 9
Terminal-Bench Hard Agents solve difficult software, system-administration and data-processing tasks in a terminal AA 62.9% #2 of 242
Agents 8 Leads on 1
AutomationBench-AA Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation AA 54.1% #31 of 133
AA-Briefcase1.1 · Elo Agents produce professional deliverables; blind pairwise comparisons produce the source Elo rating AA 1,539 #16 of 140
EnterpriseOps-Gym-AA Agents complete stateful enterprise workflows in eight domains; final database state determines success AA 51.1% #1 of 43
GDPval-AA (work tasks)2.1 Real work tasks from many professions (reports, spreadsheets, analyses), graded against expert work AA 55.5 #19 of 175
τ³-benchbanking · 1.0.1 τ-bench official · 3-banking-1.0.1: Customer-service agents converse with simulated users, retrieve domain knowledge and change state through tools Sierra 39.7% #7 of 21
EBR Bench Agents learn to play Earthborne Rangers across repeated playthroughs Epoch 39.5% #10 of 24
Furniture Assembly Models inspect furniture assembly photos, identify incorrect builds and locate the mistaken step Epoch 35.8% #20 of 31
τ³-BankingAA Agents retrieve banking knowledge and execute tools to resolve customer-support workflows AA 38.1% #21 of 134
Multimodal 2 Leads on 1
Chartography Expert chart reasoning over professional scientific, medical, financial and engineering charts Surge 34.8% #8 of 37
Arena Vision Human pairwise preferences in the stated vision vote pool Arena 1,325 #1 of 88
Writing 18
Arena Textcreative writing Human pairwise preferences in the stated text vote pool Arena 1,501 #3 of 195+14 versions & slices overall Human pairwise preferences in the stated text vote pool Arena 1,504 #4 of 196Chinese Human pairwise preferences in the stated text vote pool Arena 1,542 #7 of 181English Human pairwise preferences in the stated text vote pool Arena 1,499 #9 of 188French Human pairwise preferences in the stated text vote pool Arena 1,504 #6 of 149German Human pairwise preferences in the stated text vote pool Arena 1,487 #13 of 154instruction following Human pairwise preferences in the stated text vote pool Arena 1,502 #6 of 188Japanese Human pairwise preferences in the stated text vote pool Arena 1,506 #3 of 138Korean Human pairwise preferences in the stated text vote pool Arena 1,491 #4 of 144non-English Human pairwise preferences in the stated text vote pool Arena 1,481 #9 of 188Polish Human pairwise preferences in the stated text vote pool Arena 1,493 #8 of 133Russian Human pairwise preferences in the stated text vote pool Arena 1,503 #7 of 186Spanish Human pairwise preferences in the stated text vote pool Arena 1,496 #6 of 155overall · full cohort Human pairwise preferences in the stated text vote pool Arena 1,491 #8 of 188creative writing · full cohort Human pairwise preferences in the stated text vote pool Arena 1,494 #5 of 187
EQ-Bench Longform Creative Writing1.11 Longform Creative Writing 1.11: Models plan, revise and write a story across eight turns, graded with a writing rubric EQ-Bench 83% #4 of 95
EQ-Bench Creative Writingv3 Creative Writing v3 · Elo: Judge-rated creative writing under the v3 prompt set and pairwise ranking EQ-Bench 1,943 #10 of 91
Lech Mazur Short-Story Judges compare stories written to ten required elements; the comparison score can be negative Lech 2.9 #6 of 44
Health & law 5 Leads on 1
MLCR-AA Models synthesize long medical records; the headline credits concise answers judged both accurate and complete AA 64.4% #4 of 64
Vals LegalBench Legal reasoning and legal-language understanding on LegalBench Vals 88.6% #1 of 89
Vals MedCode Medical coding accuracy on patient encounters Vals 56.1% #4 of 78
Vals MedScribe Accuracy of clinical notes produced from consultation transcripts Vals 88.5% #11 of 80
Vals Legal Research Bench Grounded legal research; a task passes only if all criteria pass Vals 49.5% #6 of 69
Documents 3
AA-LCR (long context)1.1 Questions that need reading and connecting facts across very long documents (around 100K tokens) AA 82.3% #26 of 275
GDP.pdfAA all-pass Models reason over professional PDF documents; an attempt passes only when every atomic criterion passes 24% #12 of 133
Arena Document Human pairwise preferences in the stated document vote pool Arena 1,496 #4 of 33
Legacy · saturated or no longer reported 10
GPQA Diamond Graduate-level biology, chemistry and physics questions that can’t be solved by a quick web search 92.6% #17 of 315+1 versions & slices Epoch Models answer graduate-level science multiple-choice questions Epoch 85.9% #53 of 133
OTIS Mock AIME 20242025 Models solve the stated OTIS mock AIME mathematics problems Epoch 100% #1 of 130
SWE-rebenchall · 2026-05-15 to 2026-07-01 Repository issue resolution over the exact fresh pull-request window, under the source agent version Nebius 64.5% #1 of 6+5 versions & slices Go · 2026-05-15 to 2026-07-01 Repository issue resolution over the exact fresh pull-request window, under the source agent version Nebius 74.3% #1 of 6Java · 2026-05-15 to 2026-07-01 Repository issue resolution over the exact fresh pull-request window, under the source agent version Nebius 51.4% #3 of 6Python · 2026-05-15 to 2026-07-01 Repository issue resolution over the exact fresh pull-request window, under the source agent version Nebius 56% #1 of 6Rust · 2026-05-15 to 2026-07-01 Repository issue resolution over the exact fresh pull-request window, under the source agent version Nebius 62.4% #3 of 6TypeScript · 2026-05-15 to 2026-07-01 Repository issue resolution over the exact fresh pull-request window, under the source agent version Nebius 76.7% #2 of 6
τ²-benchAA Tests service agents that use tools while coordinating with a simulated user AA 98.5% #4 of 245
No benchmark matches
CheaperGPT-6.1 Solmax+7%-80%$0.72Remove unsupported parameters+5
OpenGLM 5.3 Flashmax-25%-99%$0.28Defaults to max · Self-hosted+4
CheaperGemini 4 Argonhigh+3%-80%*$1.99Synchronous output stops at 65,536+2
DeepSeek V4 Pro 0813maxOpen-31%-90%$0.67No image input+5
Kimi K3maxOpen-32%-70%$2.00Sampling parameters are fixed+3
Qwen3.8 27BxhighOpen-52%-95%$1.01Context: 1M → 262K+5TInkling SmallxhighOpen-55%-97%$0.45 / $1.20Context: 1M → 256K+1