Benchmark Almanac

Claude Sonnet 4.5

Claude Sonnet 4.5 ranks first on 5 of the 270 benchmarks with five or more models scored.

Anthropic Released September 29th, 2025 Family Claude 4.5 The lab's model card Compare with others
281
Benchmarks with a score
5
Ranked first, out of five or more models
42
In the top three
Anthropic Sep 2025

Claude Sonnet 4.5

Closed weights Mid-size

147.6 #80 of 256
Epoch Capabilities Index
Self-Reported Index 143.8 #145 of 404
  • Humanity's Last Exam 13.7% #25 /47
  • SWE-bench Verified 71.3% #23 /32
  • SWE-Bench Pro 43.6% #7 /24

Where it ranks

Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.

Knowledge and language

BenchmarkScoreRankSource
MMLU-Pro 87.4% #10 of 212 Official leaderboard
Phare 77.6% #4 of 62 Official leaderboard
WritingBench 80.7 #4 of 45 Official leaderboard
BullshitBench V2 79% #12 of 129 Official leaderboard
BullshitBench 65.5% #11 of 107 Official leaderboard
Slop Score 19.5 #2 of 17 Official leaderboard
AraGen 78.2% #6 of 42 Official leaderboard
ConStory-Bench 0.52 #2 of 14 Official leaderboard
IndicGenBench 60.6 #2 of 13 Official leaderboard
LMArena (Elo) 1,456 #55 of 335 Official leaderboard
MMMLU 89.1% #13 of 71 Lab model card
Judgemark v2.1 84.6 #8 of 42 Official leaderboard
EQ-Bench 3 1,535 #15 of 78 Official leaderboard
Longform Creative Writing (EQ-Bench) 71.4 #38 of 119 Official leaderboard
IFEval 85% #20 of 60 Official leaderboard
False-premise questions (Anthropic) 6.90% #2 of 6 Lab model card
Creative Writing v3 72.2 #43 of 120 Official leaderboard
DesignArena ASCII Art 1,200 #29 of 77 Official leaderboard
MultiNRC 35.8% #17 of 38 Official leaderboard
FACTS Search 69.8% #14 of 31 Official leaderboard
Humanity's Last Exam 13.7% #25 of 47 Official leaderboard
SimpleQA 24.6% #32 of 60 Lab model card
AMemGym 41.6% #6 of 11 Official leaderboard
HUMAINE 27 #28 of 50 Official leaderboard
SimpleQA Verified 29.2% #32 of 56 Official leaderboard
100Q-Hard (Anthropic) 38.3% #4 of 7 Lab model card
MultiChallenge 55.3% #26 of 42 Official leaderboard
Role-play Benchmark (MiniMax) 69.3 #5 of 8 Official leaderboard
Last Translation Benchmark 4.50 #23 of 36 Official leaderboard
FACTS Parametric 29% #29 of 45 Official leaderboard
Honesty human feedback win rate (Anthropic) 57% #4 of 6 Lab model card
Vectara Hallucination Leaderboard (HHEM) 5.50% #103 of 148 Official leaderboard
Extended NYT Connections 37.3% #78 of 103 Official leaderboard
Vectara Hallucination Leaderboard (2025 dataset) 12% #79 of 104 Official leaderboard
Document Arena 1,448 #30 of 39 Official leaderboard
Visual Aesthetic Benchmark 24.3% #14 of 18 Official leaderboard

Math and science

BenchmarkScoreRankSource
AIME 2025 100% #4 of 143 Lab model card
MATH Level 5 97.7% #5 of 94 Epoch AI run
CORE-Bench 62.2% #2 of 17 Official leaderboard
SGI-Bench 32.2% #2 of 15 Official leaderboard
GPQA Main 68.6% #11 of 60 Official leaderboard
SciPredict 22.6% #3 of 12 Official leaderboard
GPQA Diamond 82.3% #91 of 229 Epoch AI run
ForecastBench 0.13 #33 of 82 Official leaderboard
DTBench 83.2% #76 of 180 Official leaderboard
OTIS Mock AIME 2024-2025 77.8% #100 of 212 Epoch AI run
SimpleBench 54.3% #53 of 106 Official leaderboard
BRUMO 2025 90.8% #21 of 42 Official leaderboard
FrontierMath 15.2% #28 of 55 Official leaderboard
ProofBench 19% #31 of 59 Official leaderboard
ARC-AGI-2 13.6% #57 of 98 Official leaderboard
FrontierMath Tier 4 4.17% #32 of 55 Epoch AI run
ARC-AGI 63.7% #58 of 95 Official leaderboard
HMMT 2025 67.5% #33 of 53 Official leaderboard
SMT 2025 84% #25 of 40 Official leaderboard
AMO-Bench 17.6% #14 of 21 Official leaderboard
MathArena (Apex, IMO, USAMO) 1.56% #34 of 47 Official leaderboard
General365 48.6% #18 of 23 Official leaderboard
SpatialBench (LatchBio) 41.5% #11 of 14 Official leaderboard
FrontierMath Tiers 1-3 v2 23.9% #72 of 86 Official leaderboard
CMIMC 2025 66.3% #30 of 33 Official leaderboard
FrontierMath Tier 4 v2 2.44% #63 of 68 Official leaderboard

Coding

BenchmarkScoreRankSource
Spider 2.0 90.5% #2 of 23 Official leaderboard
LiveSQLBench 35.2% #4 of 36 Official leaderboard
CodeClash 1,385 #1 of 8 Official leaderboard
METR Time Horizon 117 min #4 of 31 Official leaderboard
ABC-Bench 63.2% #1 of 7 Official leaderboard
OTelBench 21.7% #3 of 14 Official leaderboard
ContextBench 58.8% #2 of 9 Official leaderboard
Snorkel Agentic Coding 57.6% #3 of 11 Official leaderboard
SWE-Bench Pro 43.6% #7 of 24 Official leaderboard
SWE-fficiency 11.6% #5 of 17 Official leaderboard
Vibe Code Bench v1 22.6% #5 of 17 Official leaderboard
Bird-SQL 66.8% #4 of 13 Official leaderboard
Time series forecasting task (Anthropic) 5.30 #3 of 9 Lab model card
SWE-Review 10.5% #2 of 5 Official leaderboard
SWE-bench-Live 36% #5 of 12 Official leaderboard
IOI (Vals) 18.3% #24 of 55 Official leaderboard
SWE-bench Verified (hard subset) 45.3% #4 of 9 Lab model card
OpenHands Index 53% #15 of 33 Official leaderboard
WeirdML 47.7% #52 of 112 Official leaderboard
GSO-Bench 14.7% #15 of 30 Official leaderboard
CooperBench 25.9% #3 of 6 Official leaderboard
SWE-bench-Live Verified 40% #1 of 2 Official leaderboard
AlgoTune 1.52× #10 of 18 Official leaderboard
DesignArena Code 1,173 #80 of 143 Official leaderboard
DesignArena Website 1,183 #85 of 151 Official leaderboard
Text-based reinforcement learning task (Anthropic) 0.85 #4 of 7 Lab model card
DesignArena 3D Design 1,140 #78 of 136 Official leaderboard
DesignArena Game Dev 1,156 #83 of 143 Official leaderboard
DesignArena UI Component 1,163 #83 of 139 Official leaderboard
VIBE (MiniMax) 85.2% #3 of 5 Official leaderboard
Vision2Web S1-2026 37.5% #3 of 5 Official leaderboard
SWE-bench Multilingual 67% #8 of 13 Official leaderboard
Quadruped reinforcement learning task (Anthropic) 1.30 #5 of 8 Lab model card
WebDev Arena 1,391 #70 of 105 WebDev Arena Leaderboard, via Epoch AI
Internal AI research evaluation suite 2 0.51 #4 of 6 Lab model card
ALE-Bench 1,192 #85 of 126 Official leaderboard
DesignArena Data Visualization 1,165 #82 of 118 Official leaderboard
DesignArena SVG 1,105 #72 of 103 Official leaderboard
Vibe Code Bench 22.6% #84 of 120 Official leaderboard
Code Arena 1,393 #90 of 128 Official leaderboard
DesignArena React Native 1,132 #40 of 56 Official leaderboard
SWE-bench Verified 71.3% #23 of 32 Epoch AI run
PostTrainBench 9.94% #18 of 23 Official leaderboard
DesignArena Agentic Overall 1,059 #58 of 74 Official leaderboard
Internal model use survey (Anthropic) 0% #4 of 5 Lab model card
APEX-SWE 31% #53 of 65 Official leaderboard
OctoCodingbench 22.8% #5 of 6 Official leaderboard
DesignArena Full Stack 1,029 #41 of 49 Official leaderboard
GameDevBench 34.8% #15 of 17 Official leaderboard
DesignArena Frontend Web App 1,041 #49 of 53 Official leaderboard

Agents and tools

BenchmarkScoreRankSource
BFCL V4 73.2% #2 of 65 Official leaderboard
AgentBench FC 58.9% #1 of 12 Official leaderboard
SnakeBench 33 #19 of 188 Official leaderboard
Terminal-Bench 46.5% #2 of 18 tbench.ai, via Epoch AI
MCP-Universe 35.1% #4 of 30 Official leaderboard
Elimination Game 5 #11 of 56 Official leaderboard
MCPMark 32.1% #6 of 28 Official leaderboard
MobileWorld 46% #4 of 18 Official leaderboard
OSWorld-MCP 49.5% #2 of 9 Official leaderboard
Step Race 3 #14 of 59 Official leaderboard
LLM-WikiRace 10.1% #5 of 21 Official leaderboard
GAIA 73.4% #16 of 64 Official leaderboard
LisanBench 2665.3 #38 of 133 Official leaderboard
CompileBench 86.7% #6 of 20 Official leaderboard
BalatroBench 11.9 #5 of 16 Official leaderboard
tau2-bench 84.7% #27 of 79 Lab model card
VitaBench 2.0 41.7% #6 of 17 Official leaderboard
OSWorld 62.9% #3 of 8 OS World Website, via Epoch AI
PinchBench 71.3% #19 of 48 Official leaderboard
EnterpriseOps-Gym 34.1% #5 of 12 Official leaderboard
Toolathlon 38.9% #21 of 50 Official leaderboard
OSWorld-Verified 62.9% #14 of 33 Official leaderboard
Terminal-Bench 2.0 46.5% #19 of 43 Official leaderboard
VITA-Bench 23.5% #8 of 18 Official leaderboard
AgencyBench 46.4% #4 of 9 Official leaderboard
DeepResearch Bench 47.5% #10 of 22 evals.futuresearch.ai, via Epoch AI
Agent Poker Bench 1,056 #8 of 17 Official leaderboard
Pencil Puzzle Bench 3.30% #25 of 53 Official leaderboard
Deep Research Bench (FutureSearch) 0.47 #10 of 21 Official leaderboard
LLM Chess 113 #61 of 123 Official leaderboard
Game Arena Coin Game 358 #20 of 39 Official leaderboard
Game Arena Poker -0.72 #17 of 31 Official leaderboard
Game Arena Clobber 226 #24 of 39 Official leaderboard
ITBench 45.8% #8 of 13 Official leaderboard
DeepPlanning 26.8% #10 of 15 Official leaderboard
τ³-bench Banking 25.3% #17 of 25 Official leaderboard
Mystery Game Puzzles 17% #58 of 84 Epoch AI run
Game Arena Word Association 127 #28 of 40 Official leaderboard
Game Arena Bargaining 161 #28 of 39 Official leaderboard
tau2-bench Verified 73.3% #8 of 11 Official leaderboard
Game Arena Ultimate Tic-Tac-Toe 176 #29 of 39 Official leaderboard
VAKRA 32.8% #3 of 4 Official leaderboard
LMArena Search Arena 1,175 #29 of 37 Official leaderboard
Game Arena Lines of Action 226 #31 of 39 Official leaderboard
Game Arena Checkers 566 #33 of 41 Official leaderboard
Game Arena Chess Openings 228 #9 of 11 Official leaderboard
Game Arena Word Art 217 #32 of 39 Official leaderboard
Kaggle Game Arena 177 #35 of 42 Official leaderboard
MCP Atlas 59.5% #31 of 37 Official leaderboard
Game Arena Nine Men's Morris 359 #33 of 39 Official leaderboard
Game Arena Reversi 150 #33 of 39 Official leaderboard
RuneBench 3.20 #65 of 76 Official leaderboard
WorldTravel 10% #6 of 7 Official leaderboard
SkillsBench 36.2% #19 of 22 Official leaderboard
Game Arena Four in a Row 157 #38 of 43 Official leaderboard
Game Arena Go 325 #35 of 39 Official leaderboard
DeepSearchQA 16% #9 of 10 Official leaderboard
Chess Puzzles 10% #85 of 93 Epoch AI run
Game Arena Dots and Boxes 544 #37 of 40 Official leaderboard
EBR-bench 2.38% #30 of 31 Epoch AI run
Game Arena Dark Hex 40 #40 of 41 Official leaderboard
Game Arena Chess 0 #42 of 42 Official leaderboard
Game Arena Five in a Row 0 #20 of 20 Official leaderboard
tau3-bench 76.4% #7 of 7 Official leaderboard
X Browse (xAI) 14.6% #4 of 4 Official leaderboard

Professional work

BenchmarkScoreRankSource
Vals Finance Agent v1 55.3% #3 of 47 Official leaderboard
LEXam 62.8% #3 of 30 Official leaderboard
PLawBench 65.9% #3 of 11 Official leaderboard
LegalBench 84.1% #39 of 138 Official leaderboard
Vals Finance Agent 54.5% #16 of 49 Official leaderboard
MedScribe 84.5% #33 of 97 Official leaderboard
CaseLaw v2 62.2% #19 of 53 Official leaderboard
TaxEval v2 73.3% #49 of 129 Official leaderboard
MedCode 44.1% #40 of 96 Official leaderboard
CorpFin v2 62% #50 of 119 Official leaderboard
Vending-Bench $2,465 #4 of 9 Official leaderboard
ProfBench 48.1% #36 of 72 Official leaderboard
Vals Index v1 53.2% #17 of 34 Official leaderboard
OccuBench 64.9% #7 of 14 Official leaderboard
Finch (FinWorkBench) 29.1% #3 of 6 Official leaderboard
MortgageTax 64% #46 of 87 Official leaderboard
FailureSensorIQ 64.2% #13 of 24 Official leaderboard
Vending-Bench 2 $3,839 #43 of 67 Official leaderboard
TaxCalcBench 31.4% #12 of 17 Official leaderboard
Vending-Bench Arena $2,602 #23 of 32 Official leaderboard
PRBench 40.8% #35 of 48 Official leaderboard
FoodTruck Bench $1,388 #45 of 59 Official leaderboard
Remote Labor Index 2.08% #15 of 19 Official leaderboard
prinzbench 20.0 #23 of 26 Official leaderboard
QF-Bench 46.2% #8 of 9 Official leaderboard
JobBench 20.7% #15 of 16 Official leaderboard

Vision, audio and robotics

BenchmarkScoreRankSource
MMMU 77.8% #14 of 136 Official leaderboard
MADQA 80.6% #2 of 18 Official leaderboard
MicroVQA 54.4% #4 of 19 Official leaderboard
MATH-Vision (MathVision) 71.1% #25 of 105 Official leaderboard
VISTA 48.8% #14 of 54 Official leaderboard
MMMU-Pro 68.9% #15 of 57 Official leaderboard
DISBench 22.1% #3 of 9 Official leaderboard
WorldVQA 20% #9 of 22 Official leaderboard
MotionBench 61.1% #5 of 12 Official leaderboard
vCX Benchmark 87% #19 of 42 Official leaderboard
OCR Arena 1,631 #6 of 11 Official leaderboard
VPCT 39.8% #16 of 27 VPCT leaderboard, via Epoch AI
VisFactor 21.5% #18 of 26 Official leaderboard
ZeroBench 5.00% #53 of 76 Official leaderboard
Math Kangaroo 2025 (MathArena) 75.8% #16 of 22 Official leaderboard
GeoBench 60% #36 of 48 Official leaderboard
MineBench 1,140 #60 of 75 Official leaderboard
ClockBench 7.80% #33 of 37 Official leaderboard
VoxelBench 1,247 #55 of 56 Official leaderboard

Safety and security

BenchmarkScoreRankSource
BioTIER Refuse 98.2% #2 of 69 Official leaderboard
AIR-Bench 2024 89.8% #3 of 82 Official leaderboard
Spiral-Bench v1.2 70.3 #2 of 21 Official leaderboard
FORTRESS 12.8% #9 of 73 Official leaderboard
Agentic misalignment suite (Anthropic) 0% #1 of 7 Lab model card
HELM Safety 97.1 #11 of 75 Official leaderboard
nocot-bench (NCRI) 94.0 #34 of 185 Official leaderboard
KORA 63.6% #8 of 42 Official leaderboard
Computer use prompt injection evaluation (Anthropic) 78% #1 of 5 Lab model card
Single-turn benign request evaluations (Anthropic) 0.02% #4 of 19 Lab model card
DTap Direct Misuse 26.6% #2 of 8 Official leaderboard
BrowseSafe-Bench 86.3% #1 of 4 Official leaderboard
Single-turn harmful request evaluations (Anthropic) 99.3% #5 of 19 Lab model card
Molecular Biology Capabilities Test 54.1% #24 of 90 Official leaderboard
AutoBaxBench 35.5% #3 of 11 Official leaderboard
SCONE-bench $436,700,000 #2 of 7 Official leaderboard
SimpleSafetyTests 100% #22 of 75 Official leaderboard
Spiral-Bench v1.0 76.4 #4 of 13 Official leaderboard
LAB-Bench CloningScenarios 66.7% #4 of 12 Lab model card
Cybench 60% #3 of 9 Claim in Opus 4.5 System Card, via Epoch AI
BABE 41.9% #2 of 6 Official leaderboard
Apollo Research strategic deception evaluations 0% #1 of 3 Lab model card
Human Pathogen Capabilities Test 56% #32 of 86 Official leaderboard
DTap Prompt Injection 25.2% #3 of 8 Official leaderboard
BinaryAudit 26.3% #8 of 20 Official leaderboard
LAB-Bench ProtocolQA 83.3% #7 of 16 Lab model card
Reward-hack-prone coding tasks (Anthropic) 1.00% #4 of 8 Lab model card
Prompt injection robustness against adaptive attackers, browser use (Anthropic) 1.90% #2 of 4 Official leaderboard
Tool use prompt injection evaluation (Anthropic) 96% #1 of 2 Lab model card
Creative biology 48.8% #4 of 7 Lab model card
LAB-Bench SeqQA 78% #7 of 12 Lab model card
Anthropic political even-handedness eval 94% #3 of 5 Official leaderboard
Average in-eval reward hacking rate (Anthropic) 15.4% #3 of 5 Lab model card
BBQ (Bias Benchmark for QA) 82.2% #16 of 25 Lab model card
CyberGym 28.9% #23 of 34 Official leaderboard
World-Class Bio 37.9% #51 of 74 Official leaderboard
Anthropic Red Team 98.2% #54 of 75 Official leaderboard
LAB-Bench FigQA 49.7% #8 of 11 Lab model card
SHADE-Arena 1.50% #13 of 16 Lab model card
UK AISI and CAISI simulated enterprise network range 9.40 #5 of 6 Official leaderboard
Gray Swan Agent Red Teaming (ART) benchmark 40% #12 of 14 Lab model card
Virology Capabilities Test (VCT) 39.7% #27 of 31 Lab model card
Impossible tasks (Anthropic) 53% #10 of 11 Lab model card
ABC Bench Liquid-Handling Robot 6.30% #32 of 34 Official leaderboard
BioTIER Permit 69.2% #67 of 69 Official leaderboard
Agentic coding malicious requests (Anthropic) 98.7% #5 of 5 Lab model card
CyberGym-E2E (615 tasks) 10.6% #4 of 4 Official leaderboard
CTI-REALM 58.7% #3 of 3 Official leaderboard
Training distribution hack rate (Anthropic) 2.00% #3 of 3 Lab model card
MCP prompt injection evaluation (Anthropic) 92% #2 of 2 Lab model card

Other

BenchmarkScoreRankSource
HELM Capabilities 76.2 #12 of 60 Official leaderboard
CAIS Risk Index 40.5 #17 of 67 Official leaderboard
ACCoRD 70.7 #51 of 152 Official leaderboard
LMCA 38.8% #60 of 147 Official leaderboard
Conceptual Reasoning Index 51.8 #64 of 151 Official leaderboard
CAIS Text Capabilities Index 22.9 #48 of 68 Official leaderboard
CAIS Vision Capabilities Index 42.1 #41 of 53 Official leaderboard
SAGE (Vals) 36.1% #64 of 81 Official leaderboard
TutorBench 49% #18 of 20 Official leaderboard

Its lab's numbers put it at 143.8, 3.8 points below its ECI

The Self-Reported Index fits one score to the 11 capability results Anthropic reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 140.2 to 147.4. How we compute it.

Its lab's numbers next to 8 leaderboards

Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.

Benchmark Lab says Board says Gap
AIME 2025 100% High school math competition AIME 2025 (python) 84.2% Leaderboard, thinking +16 pts
SWE-bench Verified 82% Agentic coding SWE-bench Verified (with parallel test-time compute) 74.8% Leaderboard, Sonar Foundation Agent +7.2 pts
OSWorld-Verified 61.4% Computer use OSWorld (official OSWorld-Verified framework, 100 max steps) 62.9% Leaderboard, 100 steps −1.5 pts
GPQA Diamond 83.4% Graduate-level reasoning GPQA Diamond 82.3% Epoch AI run, 59k thinking +1.1 pts
Vals Finance Agent 55.3% Financial analysis Finance Agent 54.5% Leaderboard, thinking +0.8 pts
tau2-bench 84.7% Mean of Agentic tool use τ2-bench Retail / Agentic tool use τ2-bench Airline / Agentic tool use τ2-bench Telecom 84.7% Leaderboard Same
CyberGym 28.9% CyberGym vulnerability reproduction, 1 trial, $2 cost constraint per trial, probability of success 0.289. 30 trials 0.667. New vulnerability discovery 0.050 at 1 trial and 0.334 at 30 trials (Figures 5.2.1.A to 5.2.1.C) 28.9% Leaderboard, Anthropic Agent Same
MMMU 77.8% Visual reasoning MMMU (validation) 77.8% Leaderboard Same

Anthropic's models over time

Anthropic's models in color, against every other model on the Epoch Capabilities Index, by release date.

50 100 150 2023 2024 2025 2026 2027 Claude Fable 5.2 Claude Haiku 5.5 Claude Sonnet 4.6 Claude Haiku 4.5 Claude 3.7 Sonnet Claude 3.5 Haiku Claude 3.5 Sonnet (October 2024) 50 100 150 2023 2024 2025 2026 2027 Claude Fable 5.2 Claude Haiku 5.5

Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)

Its launch post

The benchmarks the lab chose to headline, with the scores it gave.

Introducing Claude Sonnet 4.5 (launch post) Sep 29, 2025

Its system card

Every result in the longer documents the lab published with it. 56 of them aren't in the launch post.

System Card: Claude Sonnet 4.5 Sep 29, 2025

Hallucination

Safety

Robustness

Bias

Alignment

Bio and chem risk

Cyber risk

AI R&D risk