benchgap
Anthropic · model

Claude Haiku 4.5 Thinking benchmark scores

As of 2026-11-22, Claude Haiku 4.5 Thinking (Anthropic) has measured scores on 3 benchmarks and estimated scores on 21 more.

BenchmarkScoreSource
AA-SciCode47.1%estimated ± 3.8 pp, high confidence
AA Coding Index45.4%estimated ± 6.9 pp, medium confidence
AA-MMMU-Pro65.0%estimated ± 6.1 pp, medium confidence
SWE-bench Verified73.3%estimated ± 4.5 pp, high confidence
SWE-bench Pro53.9%estimated ± 2.9 pp, medium confidence
Mystery Game Puzzles14.5%estimated ± 4.7 pp, medium confidence
OfficeQA Pro48.5%estimated ± 7.0 pp, low confidence
ParseBench-151.2%measured
Vals SWE-bench69.1%estimated ± 7.1 pp, medium confidence
PostTrainBench v1.119.9%estimated ± 10.5 pp, low confidence
SWE-bench Multimodal0.1%estimated ± 5.6 pp, low confidence
CompileBench73.3%measured
Vibe Code Bench11.4%measured
ParseBench-261.3%estimated ± 4.3 pp, high confidence
CharXiv80.5%estimated ± 5.6 pp, low confidence
CharXiv w/o tools71.1%estimated ± 1.8 pp, low confidence
MMMU-Pro71.9%estimated ± 6.9 pp, medium confidence
PostTrainBench v119.0%estimated ± 3.1 pp, low confidence
ScreenSpot Pro74.3%estimated ± 10.1 pp, low confidence
AA LiveCodeBench85.1%estimated ± 3.4 pp, low confidence
React Native Evals57.0%estimated ± 3.0 pp, low confidence
SWE-Rebench2.2%estimated ± 6.5 pp, low confidence
PostTrainBench17.0%estimated ± 8.3 pp, low confidence
MMMU-Pro w/ Python75.2%estimated ± 2.7 pp, medium confidence