Gemini 2.5 Flash Jun 2025

Gemini 2.5 Flash ranks first on FaithJudge.

Google DeepMind Released June 17th, 2025 Compare with others

Benchmarks with a score

Ranked first, out of five or more models

In the top three

Google DeepMind Jun 2025

Gemini 2.5 Flash

Jun 2025 snapshot Closed weights

140.8 #121 of 256

Epoch Capabilities Index

Self-Reported Index 138.0 #195 of 404

Terminal-Bench 2.0 17.1% #41 /43
SimpleQA Verified 28.1% #34 /56
FrontierMath Tier 4 4.17% #30 /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
FaithJudge	6.26%	#1 of 37	Official leaderboard
RewardBench 2	77.7%	#4 of 42	Official leaderboard
SuperGPQA	59.2%	#13 of 59	Official leaderboard
Global MMLU Lite	88.4%	#4 of 16	Lab model card
LMArena (Elo)	1,409	#121 of 335	Official leaderboard
FACTS Grounding	85.3%	#6 of 15	Lab model card
ECLeKTic	36.8%	#3 of 7	Lab model card
SimpleQA	26.9%	#27 of 60	Lab model card
MRCR v2	74%	#14 of 30	Lab model card
Wer wird Millionär? LLM Benchmark	5689.0	#37 of 77	Official leaderboard
LOFT	82.1%	#4 of 8	Lab model card
FACTS Parametric	30.7%	#27 of 45	Official leaderboard
SimpleQA Verified	28.1%	#34 of 56	Official leaderboard
Humanity's Last Exam	11%	#136 of 154	Lab model card

Math and science

Benchmark	Score	Rank	Source
Misguided Attention	62.5%	#6 of 45	Official leaderboard
GPQA Diamond	82.8%	#99 of 302	Lab model card
PHYBench	33.5%	#10 of 28	Official leaderboard
FrontierMath Tier 4	4.17%	#30 of 55	Epoch AI run
DTBench	76.5%	#102 of 180	Official leaderboard
FrontierMath	4.84%	#46 of 70	Epoch AI run
AIME 2025	72%	#109 of 143	Lab model card
HiddenMath-Hard	75.5%	#2 of 2	Lab model card

Coding

Benchmark	Score	Rank	Source
LiveCodeBench v5	63.9%	#19 of 47	Lab model card
AutoCodeBench	45.7%	#14 of 34	Official leaderboard
Aider Polyglot	61.9%	#13 of 30	Lab model card
WeirdML	41%	#74 of 112	Official leaderboard
SWE-bench Verified	28.7%	#49 of 60	Official leaderboard
IOI (Vals)	2.61%	#48 of 55	Official leaderboard

Agents and tools

Benchmark	Score	Rank	Source
Step Race	4	#6 of 59	Official leaderboard
Elimination Game	5	#17 of 56	Official leaderboard
LisanBench	362.3	#76 of 133	Official leaderboard
DeepSearchQA	25.9%	#6 of 10	Official leaderboard
Galileo Agent Leaderboard v2	38%	#14 of 21	Official leaderboard
GAIA	30.2%	#44 of 64	Official leaderboard
Terminal-Bench	17.1%	#16 of 18	tbench.ai, via Epoch AI
Terminal-Bench 2.0	17.1%	#41 of 43	Official leaderboard

Professional work

Benchmark	Score	Rank	Source
ProfBench	49.2%	#31 of 72	Official leaderboard
MedScribe	83%	#45 of 97	Official leaderboard
MedCode	40.4%	#62 of 96	Official leaderboard
PRBench	41%	#33 of 48	Official leaderboard

Vision, audio and robotics

Benchmark	Score	Rank	Source
Math-VR	60.5%	#4 of 26	Official leaderboard
MMMU	79.7%	#23 of 103	Lab model card
ActivityNet-QA	65.1%	#2 of 8	Lab model card
MADQA	71.8%	#5 of 18	Official leaderboard
Perception Test MCVQA	75.1%	#2 of 6	Lab model card
SpatialViz-Bench	36.9%	#3 of 8	Claim in arxiv.org, via Epoch AI
Vibe-Eval	65.4%	#4 of 9	Lab model card
Audio MC	40%	#9 of 18	Official leaderboard
YouCook2	177.6	#2 of 4	Lab model card
Vision Arena	1,215	#69 of 136	Official leaderboard
Video-MME	81.5%	#26 of 51	Lab model card
VATEX	65.2	#3 of 5	Lab model card
BetterChartQA	67.3%	#2 of 3	Lab model card
Video-MMMU	79.2%	#25 of 36	Lab model card
LVBench	62.7%	#24 of 31	Lab model card
CoVoST2	36.1	#11 of 14	Lab model card
ZeroBench	2.00%	#30 of 31	Lab model card
MINERVA	60.7%	#6 of 6	Lab model card
EgoTempo	36.7%	#5 of 5	Lab model card
QVHighlights	52.4%	#5 of 5	Lab model card
VATEX ZH	43.9	#4 of 4	Lab model card
1H-VideoQA	67.5%	#2 of 2	Lab model card
Neptune	84.3%	#2 of 2	Lab model card

Safety and security

Benchmark	Score	Rank	Source
Hack the Box challenge set (Google DeepMind)	15.4%	#1 of 4	Lab model card
InterCode-CTF	98.7%	#2 of 4	Lab model card
Divergence attack (extractable memorization)	59%	#1 of 2	Lab model card
Indirect prompt injection, Actor Critic attack	40.8%	#1 of 2	Lab model card
Indirect prompt injection, Beam Search attack	4.20%	#1 of 2	Lab model card
WMDP Chemistry	75%	#5 of 7	Lab model card
Human Pathogen Capabilities Test v2 (HPCT v2)	43.5%	#52 of 71	Official leaderboard
In-house CTF suite (Phuong et al. 2024)	76.9%	#3 of 4	Lab model card
Lakera B3 (Backbone Breaker Benchmark)	49.8%	#27 of 35	Official leaderboard
LAB-Bench CloningScenarios	44%	#10 of 12	Lab model card
Molecular Biology Capabilities Test	39.8%	#76 of 90	Official leaderboard
WMDP Biology	84%	#7 of 8	Lab model card
LAB-Bench ProtocolQA	64%	#16 of 16	Lab model card
LAB-Bench SeqQA	58%	#12 of 12	Lab model card
Cyber key skills benchmark v1	8.30%	#4 of 4	Lab model card
SecureBio VMQA, single choice	48%	#4 of 4	Lab model card
Dangerous Content policy violations (from ART)	26.9%	#2 of 2	Lab model card
Helpfulness violations (from ART)	6.60%	#2 of 2	Lab model card
Indirect prompt injection, TAP attack	53.6%	#2 of 2	Lab model card

Other

Benchmark	Score	Rank	Source
SAGE (Vals)	44.8%	#44 of 81	Official leaderboard
Conceptual Reasoning Index	37.7	#104 of 151	Official leaderboard
CAIS Vision Capabilities Index	41.1	#42 of 53	Official leaderboard
ACCoRD	45.4	#122 of 152	Official leaderboard
CAIS Text Capabilities Index	13.2	#60 of 68	Official leaderboard
CAIS Risk Index	70.0	#65 of 67	Official leaderboard

Its lab's numbers put it at 138.0, 2.8 points below its ECI

The Self-Reported Index fits one score to the 18 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 135.5 to 140.5. How we compute it.

Its lab's numbers next to 3 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	What differs
SWE-bench Verified	60.4% SWE-Bench Verified	28.7% Leaderboard, mini-SWE-agent	+32 pts	Google uses its internal coding agent, while the board uses mini-SWE-agent.
ECLeKTic	36.8% ECLeKTic (Table 3)	30.9% Leaderboard	+5.9 pts
FACTS Grounding	85.3% FACTS Grounding	86.4% Leaderboard	−1.1 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.

Gemini 2.5 Flash Model Card, GA Thinking column (Google DeepMind) Jun 17, 2025

Humanity's Last Exam (no tools), GA Thinking 11
GPQA diamond single attempt 82.8
AIME 2025 single attempt 72
LiveCodeBench v5 single attempt 63.9
Aider Polyglot whole 61.9 / diff 56.7 61.9
SWE-Bench Verified 60.4
SimpleQA 26.9
FACTS Grounding 85.3
MMMU single attempt 79.7
Vibe-Eval (Reka) 65.4
MRCR v2 128k (average), earlier non-8-needle version (1M pointwise 32) 74
Global MMLU (Lite) 88.4

Its system card

Every result in the longer documents the lab published with it. 45 of them aren't in the launch post.

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities Jul 7, 2025