Gemini 2.5 Flash Sep 2025

Gemini 2.5 Flash ranks first on JointAVBench.

Google DeepMind Released September 25th, 2025 Compare with others

Benchmarks with a score

Ranked first, out of five or more models

In the top three

Google DeepMind Sep 2025

Gemini 2.5 Flash

Sep 2025 snapshot Closed weights

141.9 #111 of 256

Epoch Capabilities Index

Self-Reported Index 136.4 #213 of 404

Toolathlon 3.70% #50 /50
ZeroBench 3.00% #61 /76
FrontierMath 4.84% #40 /55

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

Benchmark	Score	Rank	Source
HELM Arabic	82.5	#2 of 28	Official leaderboard
IFEval	89.8%	#10 of 60	Official leaderboard
FACTS Grounding	87.5%	#3 of 15	Lab model card
IndicGenBench	56.9	#4 of 13	Official leaderboard
Vectara Hallucination Leaderboard (2025 dataset)	7.80%	#39 of 104	Official leaderboard
LMArena (Elo)	1,403	#126 of 335	Official leaderboard
HELM Arabic Enterprise	68.2	#9 of 23	Official leaderboard
Global MMLU Lite	91.5%	#19 of 43	Official leaderboard
SimpleQA	25.3%	#30 of 60	Lab model card
EQ-Bench 3	1,178	#45 of 78	Official leaderboard
Phare	63.6%	#41 of 62	Official leaderboard
Judgemark v2.1	63.0	#30 of 42	Official leaderboard
Longform Creative Writing (EQ-Bench)	48.6	#86 of 119	Official leaderboard
Creative Writing v3	58.7	#87 of 120	Official leaderboard
MRCR v2	52.4%	#22 of 30	Lab model card
BullshitBench V2	19%	#96 of 129	Official leaderboard
BullshitBench	14.6%	#84 of 107	Official leaderboard
Humanity's Last Exam	13.2%	#129 of 154	Lab model card
FACTS Search	60%	#26 of 31	Official leaderboard
SimpleQA Verified	27.8%	#20 of 22	Lab model card
Slop Score	77.7	#16 of 17	Official leaderboard

Math and science

Benchmark	Score	Rank	Source
HiPhO	61.5%	#11 of 36	Official leaderboard
GPQA Diamond	80.8%	#114 of 302	Lab model card
ForecastBench	0.14	#43 of 82	Official leaderboard
AMO-Bench	18.1%	#13 of 21	Official leaderboard
HMMT 2025	64.2%	#35 of 53	Official leaderboard
SimpleBench	41.2%	#72 of 106	Official leaderboard
AIME 2025	75.6%	#102 of 143	Lab model card
FrontierMath	4.84%	#40 of 55	Official leaderboard
miniCTX-v2	26.6%	#3 of 4	Official leaderboard
GPQA Main	39%	#47 of 60	Official leaderboard
SciArena	933	#27 of 34	Official leaderboard
BRUMO 2025	83.3%	#34 of 42	Official leaderboard
SMT 2025	75.5%	#34 of 40	Official leaderboard
CMIMC 2025	54.4%	#33 of 33	Official leaderboard
General365	39.6%	#23 of 23	Official leaderboard

Coding

Benchmark	Score	Rank	Source
LiveCodeBench v5	71.7%	#7 of 47	Lab model card
LiveSQLBench	23.8%	#20 of 36	Official leaderboard
WeirdML	41.9%	#70 of 112	Official leaderboard
DesignArena Website	1,108	#113 of 151	Official leaderboard
DesignArena UI Component	1,084	#108 of 139	Official leaderboard
DesignArena Code	1,094	#113 of 143	Official leaderboard
SWE-bench Verified	54%	#123 of 155	Lab model card
DesignArena 3D Design	1,054	#108 of 136	Official leaderboard
DesignArena Game Dev	1,056	#117 of 143	Official leaderboard
DesignArena Data Visualization	1,132	#97 of 118	Official leaderboard
SWE-fficiency	4.06%	#14 of 17	Official leaderboard
DesignArena SVG	1,010	#91 of 103	Official leaderboard

Agents and tools

Benchmark	Score	Rank	Source
SnakeBench	31	#34 of 188	Official leaderboard
BFCL V4	56.2%	#12 of 65	Official leaderboard
GAIA	58.8%	#29 of 64	Official leaderboard
ITBench	51.7%	#6 of 13	Official leaderboard
MCP-Universe	21.6%	#23 of 30	Official leaderboard
MCPMark	9.06%	#23 of 28	Official leaderboard
PinchBench	50.6%	#42 of 48	Official leaderboard
Terminal-Bench	16.9%	#17 of 18	Terminal-Bench v2 Leaderboard, via Epoch AI
Toolathlon	3.70%	#50 of 50	Official leaderboard
CompileBench	60%	#20 of 20	Official leaderboard

Professional work

Benchmark	Score	Rank	Source
FailureSensorIQ	65.5%	#9 of 24	Official leaderboard
TaxEval v2	72.7%	#57 of 129	Official leaderboard
LegalBench	82.6%	#66 of 138	Official leaderboard
Vals Finance Agent v1	31.2%	#23 of 47	Official leaderboard
MortgageTax	62.6%	#50 of 87	Official leaderboard
CorpFin v2	59.8%	#71 of 119	Official leaderboard
MedScribe	78.5%	#60 of 97	Official leaderboard
MedCode	40.5%	#60 of 96	Official leaderboard
Vending-Bench 2	$549	#56 of 67	Official leaderboard
Vending-Bench Arena	$281	#31 of 32	Official leaderboard

Vision, audio and robotics

Benchmark	Score	Rank	Source
JointAVBench	58%	#1 of 6	Official leaderboard
MMMU	80.3%	#21 of 103	Lab model card
VisualPuzzles	44.6%	#10 of 29	Official leaderboard
OCR Arena	1,709	#4 of 11	Official leaderboard
VPCT	46.2%	#10 of 27	VPCT leaderboard, via Epoch AI
Vision Arena	1,229	#62 of 136	Official leaderboard
AHELM	49.7	#5 of 10	Official leaderboard
DECIDE	36.6	#4 of 8	Official leaderboard
RoboRewardBench	0.94	#13 of 22	Official leaderboard
Blueprint-Bench	0.36	#5 of 8	Official leaderboard
Vibe-Eval	62%	#6 of 9	Lab model card
ClockBench	11.1%	#27 of 37	Official leaderboard
ZeroBench	3.00%	#61 of 76	Official leaderboard

Safety and security

Benchmark	Score	Rank	Source
AIR-Bench 2024	68.7%	#35 of 82	Official leaderboard
HELM Safety	91.2	#39 of 75	Official leaderboard
KORA	36.4%	#24 of 42	Official leaderboard
Anthropic Red Team	98.8%	#43 of 75	Official leaderboard
Spiral-Bench v1.0	49.1	#8 of 13	Official leaderboard
Spiral-Bench v1.2	36.5	#13 of 21	Official leaderboard
SimpleSafetyTests	98%	#48 of 75	Official leaderboard

Other

Benchmark	Score		Rank	Source
SAGE (Vals)	44.1%		#46 of 81	Official leaderboard
HELM Capabilities	62.6		#37 of 60	Official leaderboard

Its lab's numbers put it at 136.4, 5.5 points below its ECI

The Self-Reported Index fits one score to the 12 capability results Google DeepMind 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 134.2 to 138.5. How we compute it.

Its lab's numbers next to 2 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	75.6% AIME 2025 single attempt, model card	70.8% Leaderboard, thinking	+4.8 pts
Global MMLU Lite	87.9% Global MMLU (Lite), model card	91.5% Leaderboard	−3.5 pts

Google DeepMind's models over time

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

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.

Continuing to bring you our latest models, with an improved Gemini 2.5 Flash and Flash-Lite release (Google Developers Blog) Sep 25, 2025

Its system card

Every result in the longer documents the lab published with it. 5 of them aren't in the launch post.