benchgap
Anthropic · model

Claude Sonnet 4.5 Thinking benchmark scores

As of 2026-11-22, Claude Sonnet 4.5 Thinking (Anthropic) has measured scores on 9 benchmarks and estimated scores on 45 more.

BenchmarkScoreSource
AA-GPQA Diamond85.5%estimated ± 2.7 pp, high confidence
AA-HLE29.6%estimated ± 6.0 pp, medium confidence
AA-SciCode48.9%estimated ± 3.8 pp, high confidence
Artificial Analysis Intelligence Index19.6%estimated ± 5.8 pp, medium confidence
CritPt8.7%estimated ± 5.1 pp, medium confidence
GPQA Diamond85.1%estimated ± 1.5 pp, medium confidence
HLE43.8%estimated ± 6.1 pp, medium confidence
LiveCodeBench v683.2%estimated ± 2.2 pp, low confidence
AA-Omniscience Accuracy40.9%estimated ± 6.5 pp, medium confidence
SciCode47.2%estimated ± 4.5 pp, medium confidence
AA Coding Index52.5%estimated ± 5.6 pp, medium confidence
SWE-bench Verified76.7%estimated ± 4.4 pp, high confidence
SWE-bench Pro54.8%estimated ± 2.9 pp, medium confidence
MMLU-Pro86.2%estimated ± 4.8 pp, high confidence
FailureSensorIQ64.2%estimated ± 1.7 pp, high confidence
GPQA85.0%estimated ± 1.6 pp, high confidence
MMLU94.3%estimated ± 1.2 pp, medium confidence
Vals GPQA Diamond81.6%measured
Vals LiveCodeBench73.0%measured
Vals MMLU-Pro87.4%measured
Chess Puzzles19.5%estimated ± 12.3 pp, low confidence
AA MMLU-Pro87.9%estimated ± 1.1 pp, high confidence
ARC-AGI-163.7%measured
ARC-AGI-213.6%measured
CursorBench 3.143.9%estimated ± 5.8 pp, low confidence
CursorBench 3.252.7%estimated ± 5.4 pp, low confidence
FrontierCode 1.1 Main23.8%estimated ± 5.2 pp, low confidence
FrontierSWE v26.4%estimated ± 9.5 pp, low confidence
Mystery Game Puzzles19.1%estimated ± 4.7 pp, medium confidence
NYT Connections37.3%measured
Vals SWE-bench70.0%measured
VulcanBench v378.6%estimated ± 5.6 pp, low confidence
CursorBench 4.030.8%estimated ± 3.4 pp, medium confidence
DeepSWE56.3%estimated ± 7.2 pp, low confidence
HLE w/o tools39.7%estimated ± 4.9 pp, high confidence
PostTrainBench v1.115.4%estimated ± 3.7 pp, medium confidence
StealthMark41.3%estimated ± 10.7 pp, low confidence
ARC-AGI-30.4%estimated ± 8.6 pp, low confidence
SkillsBench official leaderboard (BenchFlow)43.8%estimated ± 6.1 pp, medium confidence
BinaryAudit35.1%estimated ± 12.1 pp, low confidence
CompileBench84.4%measured
OTelBench22.3%estimated ± 6.6 pp, medium confidence
Vibe Code Bench22.6%measured
FACTS Multimodal22.7%estimated ± 4.0 pp, medium confidence
NL2Repo40.1%estimated ± 4.7 pp, high confidence
PostTrainBench v119.0%estimated ± 3.1 pp, medium confidence
AA LiveCodeBench86.6%estimated ± 2.6 pp, high confidence
React Native Evals70.7%estimated ± 3.0 pp, low confidence
SWE-Rebench42.8%estimated ± 6.5 pp, low confidence
PostTrainBench12.9%estimated ± 3.7 pp, medium confidence
LABBench281.7%estimated ± 1.8 pp, low confidence
ScarfBench41.2%estimated ± 12.3 pp, low confidence
OpenHarmony Bench v1.051.1%estimated ± 3.2 pp, high confidence
BioBasicsBench34.3%estimated ± 5.1 pp, low confidence