Updated daily at 05:00 UTC

Vision-language

Gemini 3 Flash

By Google DeepMind, in the Gemini Flash family.

Released: 17 Dec 2025
Status: Legacy
Context: 1M
Max output: 65.5K
Input per 1M tokens: $0.50
Output per 1M tokens: $3.00
Open weights: No
License: proprietary
Parameters: —
Knowledge cutoff: Jan 2025
Reasoning: Yes
Modalities: text, image, video, audio, pdf → text

Category scores

Out of 100, from independent and third-party results only, by method category-2026-09-29.1.

Category scores
Category	Score	Rank	Coverage
Knowledge & reasoning	47.0	23	3 of 4
Coding & software engineering	9.8	Not ranked	1 of 5
Agentic tool use	49.0	4	2 of 2
Vision & multimodal understanding	73.4	2	2 of 3
Factuality & hallucination	66.8	9	1 of 1
Human preference	51.0	26	4 of 5

Benchmark results

Every result we hold for this model, with who ran it and where it was published.

Knowledge & reasoning

Knowledge & reasoning results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
Chess Puzzles	Puzzles solved	40.0%	Independent	high	6 Aug 2026	Epoch AI
Chess Puzzles	Puzzles solved	38.0%	Independent	—	17 Dec 2025	Epoch AI
DTBench (Decision Theory Benchmark)	Accuracy (capability questions)	89.1%	Third party	high	—	https://conceptualreasoning.ai/dtbench via Epoch AI
LMCA (Language Model Conceptual Argumentation)	Correlation with expert ratings (Pearson r)	0.4313	Third party	high	—	https://conceptualreasoning.ai/lmca via Epoch AI
Mystery Game Puzzles	Puzzles solved	26.0%	Independent	low	5 Aug 2026	Epoch AI
Mystery Game Puzzles	Puzzles solved	25.0%	Independent	minimal	5 Aug 2026	Epoch AI
Mystery Game Puzzles	Puzzles solved	20.0%	Independent	high	5 Aug 2026	Epoch AI
SimpleBench	Score (AVG@5)	61.1%	Third party	—	—	https://lmcouncil.ai/benchmarks via Epoch AI

Scientific reasoning

Scientific reasoning results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
GPQA Diamond	Accuracy	89.4%	Independent	high	6 Aug 2026	Epoch AI
GPQA Diamond	Accuracy	83.2%	Independent	—	17 Dec 2025	Epoch AI

Mathematics

Mathematics results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
FrontierMath Tier 4	Accuracy	17.1%	Independent	—	11 Jun 2026	Epoch AI
FrontierMath Tier 4 (v1)	Accuracy	4.2%	Independent	—	17 Dec 2025	Epoch AI
FrontierMath Tiers 1-3	Accuracy	51.2%	Independent	—	11 Jun 2026	Epoch AI
FrontierMath Tiers 1-3 (v1)	Accuracy	35.6%	Independent	—	17 Dec 2025	Epoch AI
OTIS Mock AIME 2024-2025	Accuracy	95.6%	Independent	high	6 Aug 2026	Epoch AI
OTIS Mock AIME 2024-2025	Accuracy	92.8%	Independent	—	17 Dec 2025	Epoch AI

Coding & software engineering

Coding & software engineering results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
GSO	OPT@1	9.8%	Third party	—	—	https://gso-bench.github.io/index.html via Epoch AI
SWE-bench Verified	% resolved	75.4%	Independent	—	18 Feb 2026	Epoch AI
Terminal-Bench 2.0	Accuracy (mean over trials)	64.3%	Third party	—	—	https://www.tbench.ai/leaderboard/terminal-bench/2.0 via Epoch AI
Terminal-Bench 2.0	Accuracy (mean over trials)	51.7%	Third party	—	—	https://www.tbench.ai/leaderboard/terminal-bench/2.0 via Epoch AI
Terminal-Bench 2.0	Accuracy (mean over trials)	51.0%	Third party	—	—	Terminal-Bench v2 Leaderboard via Epoch AI
Terminal-Bench 2.0	Accuracy (mean over trials)	47.4%	Third party	—	—	https://www.tbench.ai/leaderboard/terminal-bench/2.0 via Epoch AI
WeirdML (v2)	Average accuracy	61.6%	Third party	—	—	https://htihle.github.io/weirdml.html via Epoch AI

Agentic tool use

Agentic tool use results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
BALROG	Average progress	48.1%	Third party	—	—	https://balrogai.com/ via Epoch AI
Deep Research Bench (FutureSearch)	Average score	49.8%	Third party	low	—	https://evals.futuresearch.ai/#drb via Epoch AI
Deep Research Bench (FutureSearch)	Average score	49.0%	Third party	minimal	—	https://evals.futuresearch.ai/#drb via Epoch AI
Deep Research Bench (FutureSearch)	Average score	47.9%	Third party	high	—	https://evals.futuresearch.ai/#drb via Epoch AI

Vision & multimodal understanding

Vision & multimodal understanding results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
GeoBench	Country accuracy	88.0%	Third party	—	—	GeoBench leaderboard via Epoch AI
Visual Physics Comprehension Test (VPCT)	Correct	72.6%	Third party	—	—	VPCT leaderboard via Epoch AI

Factuality & hallucination

Factuality & hallucination results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
SimpleQA Verified	Accuracy (% correct)	66.8%	Independent	high	27 Aug 2026	Epoch AI

Human preference

Human preference results
Benchmark	Metric	Value	Provenance	Setting	Evaluated	Source
Code Arena: WebDev (arena.ai)	Arena score (Elo scale)	1439	Independent	—	25 Sep 2026	Arena
Code Arena: WebDev (arena.ai)	Votes	10900	Independent	—	25 Sep 2026	Arena
Document Arena (arena.ai)	Arena score (Elo scale)	1439	Independent	—	13 Sep 2026	Arena
Document Arena (arena.ai)	Votes	7158	Independent	—	13 Sep 2026	Arena
Text Arena (arena.ai)	Arena score (Elo scale)	1473	Independent	—	25 Sep 2026	Arena
Text Arena (arena.ai)	Votes	31283	Independent	—	25 Sep 2026	Arena
Vision Arena (arena.ai)	Arena score (Elo scale)	1273	Independent	—	28 Sep 2026	Arena
Vision Arena (arena.ai)	Votes	39181	Independent	—	28 Sep 2026	Arena

Other Gemini Flash versions

Sources