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 20 #123 of 339
Knowledge 4
AA OmniscienceAccuracy Measures correct answers on the AA Omniscience factual knowledge test AA 21.9% #138 of 323+1 versions & slices Hallucination Rate Measures how often attempted answers are incorrect on the AA Omniscience factual knowledge test AA 51.1% #94 of 323
AA Omniscience Index Scores factual knowledge while penalizing incorrect answers to encourage admitting uncertainty AA -18 #121 of 323
IFBench (instructions) Checks whether the model follows precise formatting and content instructions AA 64.4% #72 of 248
Reasoning 4
ARC-AGI-1public evaluation Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 43.4% #54 of 64+1 versions & slices semi-private Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 33% #56 of 67
ARC-AGI-2public evaluation Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 3.5% #54 of 66+1 versions & slices semi-private Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit ARC 1.9% #59 of 69
Science 3
Humanity's Last Exam Very hard expert questions across many fields, written to stump today’s models AA 15.9% #141 of 339
CritPt (physics) Research-level physics problems that need many careful reasoning steps AA 5.1% #80 of 326
SciCode Writing code for real scientific computing problems (physics, biology, math) AA 47.2% #66 of 142
Coding 4
Terminal-Bench4.0 Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite 0.5% #65 of 134+1 versions & slices 2.1 Older, easier version of Terminal-Bench; Still widely reported AA 60.7% #53 of 156
ITBench SRE IT incident tasks: diagnose and fix broken systems like a site-reliability engineer AA 24.4% #33 of 37
Terminal-Bench Hard Agents solve difficult software, system-administration and data-processing tasks in a terminal AA 33.3% #50 of 242
Agents 6
AutomationBench-AA Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation AA 6% #83 of 135
AA-Briefcase1.1 · Elo Agents produce professional deliverables; blind pairwise comparisons produce the source Elo rating AA 673 #88 of 142
APEX-Agents Agents complete long-horizon professional-services tasks across realistic applications AA 24.9% #17 of 29
EnterpriseOps-Gym-AA Agents complete stateful enterprise workflows in eight domains; final database state determines success AA 31.5% #33 of 43
GDPval-AA (work tasks)2.1 Real work tasks from many professions (reports, spreadsheets, analyses), graded against expert work AA 947 #78 of 177
τ³-BankingAA Agents retrieve banking knowledge and execute tools to resolve customer-support workflows AA 27.4% #39 of 134
Multimodal 2
MMMU-Pro (vision) College-level questions that require understanding images: charts, diagrams, photos AA 59.5% #99 of 128
Arena Vision Human pairwise preferences in the stated vision vote pool Arena 1,196 #66 of 88
Writing 18
Arena Textcreative writing · October 22, 2026 Human pairwise preferences in the stated text vote pool Arena 1,314 #119 of 197+16 versions & slices creative writing Human pairwise preferences in the stated text vote pool Arena 1,337 #109 of 195overall Human pairwise preferences in the stated text vote pool Arena 1,401 #98 of 196Chinese Human pairwise preferences in the stated text vote pool Arena 1,391 #118 of 198English Human pairwise preferences in the stated text vote pool Arena 1,381 #117 of 205French Human pairwise preferences in the stated text vote pool Arena 1,395 #103 of 164German Human pairwise preferences in the stated text vote pool Arena 1,368 #94 of 167instruction following Human pairwise preferences in the stated text vote pool Arena 1,362 #110 of 205Japanese Human pairwise preferences in the stated text vote pool Arena 1,343 #81 of 152Korean Human pairwise preferences in the stated text vote pool Arena 1,320 #95 of 156non-English Human pairwise preferences in the stated text vote pool Arena 1,360 #112 of 205Polish Human pairwise preferences in the stated text vote pool Arena 1,366 #93 of 142Russian Human pairwise preferences in the stated text vote pool Arena 1,362 #111 of 203Spanish Human pairwise preferences in the stated text vote pool Arena 1,371 #112 of 168overall · October 15, 2026 Human pairwise preferences in the stated text vote pool Arena 1,372 #105 of 188creative writing · October 15, 2026 Human pairwise preferences in the stated text vote pool Arena 1,314 #115 of 187overall · October 22, 2026 Human pairwise preferences in the stated text vote pool Arena 1,372 #109 of 198
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 57.8% #55 of 96
Health & law 5
MLCR-AA Models synthesize long medical records; the headline credits concise answers judged both accurate and complete AA 16.1% #34 of 103
Vals LegalBench Legal reasoning and legal-language understanding on LegalBench Vals 77.9% #71 of 89
Vals MedCode Medical coding accuracy on patient encounters Vals 41% #48 of 78
Vals MedScribe Accuracy of clinical notes produced from consultation transcripts Vals 77.1% #55 of 80
Vals Legal Research Bench Grounded legal research; a task passes only if all criteria pass Vals 6.3% #68 of 73
Documents 2
AA-LCR (long context)1.1 Questions that need reading and connecting facts across very long documents (around 100K tokens) AA 67.3% #137 of 317
GDP.pdfAA all-pass Models reason over professional PDF documents; an attempt passes only when every atomic criterion passes 7.8% #72 of 135
Legacy · saturated or no longer reported 2
GPQA Diamond Graduate-level biology, chemistry and physics questions that can’t be solved by a quick web search 76.1% #143 of 315
τ²-benchAA Tests service agents that use tools while coordinating with a simulated user AA 52.6% #110 of 245
No benchmark matches
CheaperClaude Haiku 5.5xhigh×4.4-57%$0.12Remove sampling parameters+5
OpenMiMo v2.6 Prothinking×5.2+18%$0.13Pass back reasoning_content+3
CheaperMiniMax M3.1 Flashthinking×4.3-43%$0.22No strict JSON+1
CheaperAgnes 2.5 Pro Beta×3.1-68%$0.10 / $0.30
GLM 5.3 FlashmaxOpen×3-49%$0.28Defaults to max · Self-hosted+4
DeepSeek V4.1 Flash 1017maxOpen×4.7+14%$0.26Pass back reasoning_content · Chat Completions+4
Kimi K3maxOpen×4.6+1,197%$2.00Sampling parameters are fixed+3
Step 5 PreviewthinkingOpen×4.1+208%$1.03web_search not available · Chat Completions+2