Atlas

Models

← All models

Gemini 2.5 Pro Experimental 03-25

Google DeepMind · 2025-03-25

The March 25, 2025 experimental checkpoint of Gemini 2.5 Pro, kept distinct from preview aliases and the stable June release.

Benchmark scores

BenchmarkScore
Aider Polyglot68.6
AIME 2024 (avg@8)87.5
AIME 2024-202585.8
AIME 202586.7
AIME 2025 (avg@8)84.2
BALROG43.3
BigCodeBench-Hard Complete36.5
BigCodeBench-Hard Instruct29.7
Capability146.0
CorpFin v259.8
CorpFin v2 - Exact Pages59.9
CorpFin v2 - Max Fitting Context57.8
CorpFin v2 - Shared Max Context61.8
Creative Writing v286.2
Creative Writing v3 Elo1390
Creative Writing v3 Rubric Score16.0
EnigmaEval4.1
Fiction.liveBench66.7
GeoBench ACW Country Accuracy81.0
Global MMLU Lite89.8
GPQA Diamond80.3
GPQA Diamond (Vals protocol average)80.8
HardGeoBench <=100mi50.0
HardGeoBench <=10mi35.0
HardGeoBench Avg Distance1172
HardGeoBench Country Accuracy56.3
LegalBench84.3
LegalBench - Conclusion Tasks88.3
LegalBench - Interpretation Tasks82.7
LegalBench - Issue Tasks87.1
Loading Atlas data…