whatmodel.sh / Gemini 3 Flash (Preview)

Gemini 3 Flash (Preview)

+ contribute

Gemini 3 Flash (Preview) is a closed-weights language model from Google, released December 2025.

Catalogue identity notes

Comparison-only. API model 'gemini-3-flash-preview' (MathArena config gemini/gemini-3-flash, date 2025-12-17, thinking_level high; ARC Prize folders gemini-3-flash-preview-thinking-*). Distinct from gemini-3-5/3-6/3-7/3-8-flash.

72.3±0.7#63 of 719 LLMsfrom 28 benchmark results across 12 sources

$ how it compares

The LLMs ranked either side of Gemini 3 Flash (Preview), out of 828 with a score, then more from Google.

#	model	score	where	RAM	released
66	Qwen3.6 Max Preview Alibaba	72±1	api $2.92/M		2026-04
67	Claude Sonnet 4.6 Anthropic	72±1	api $6.00/M		2026-02
68	Gemini 3 Flash (Preview) Google	72±1	api $1.13/M		2025-12
69	DeepSeek V4 Pro (Preview, Apr 2026) DeepSeek	72±1	local	~604GB	2026-04
70	Claude Opus 4.5 Anthropic	72±1	api $10.00/M		2025-11
	more from Google
9	Gemini 4 Argon (limited preview) Google	82±1	limited preview		2026-09
23	Gemini 3.8 Flash Google	77±1	api $1.50/M		2026-09
32	Gemini 3.7 Flash Google	76±1	api $1.50/M		2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #78 of 300 · 17% of its score · 6 tests

Beats GPT-5 here, though they rank higher overall. Trails MiniMax-M2.7, MiMo-V2.5 (310B-A15B), Qwen3.6 27B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified	75.4% ✎	▁··▁·▁▂▁▄█▄▁ beats 19/29	hard · avg 19%	▇▇▇ 7.6%	sharp · at its level
τ³-bench banking_knowledge (v1.0.1 grading)	27.3% ✎	▃▃▂▃·▆▃▂█▅▅▂ beats 12/25	very hard · avg 4.1%	▇ 2.7%	at its level
Vibe Code Bench (test split, 50 web apps)	20.2% ✎	▃··▃·▃██▃▃▃█ beats 7/14	very hard · avg 2.3%	▇ 2.2%	sharp · at its level · noisy
τ³-bench airline (tau2-bench ≥ v1.0.1)	82.5% ✎	▃······▃·▃█▃ beats 4/6	medium · avg 49%	▇ 1.9%	typical
τ³-bench retail (tau2-bench ≥ v1.0.1)	76.8% ✎	▄▄·█··▄·▄··▄ beats 3/6	medium · avg 46%	▇ 1.6%	typical
τ³-bench telecom (tau2-bench ≥ v1.0.1)	91.2% ✎	█·▄█▄······▄ beats 4/6	medium · avg 62%	▇ 1.3%	typical

coding▁▁▁▃▆██▇▆▄▃▁ #68 of 458 · 17% of its score · 3 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-5.2, GPT-5 here, though they rank higher overall. Trails Gemini 3.5 Flash-Lite, MiMo-V2.5 (310B-A15B), MiMo-V2.5-Pro (1T-A42B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified	75.4% ✎	▁··▁·▁▂▁▄█▄▁ beats 19/29	hard · avg 19%	▇▇▇ 7.6%	sharp · at its level
Arena WebDev (Code Arena)	1438.9 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 57/119	avg 981	▇▇▇ 7.1%	sharp
Vibe Code Bench (test split, 50 web apps)	20.2% ✎	▃··▃·▃██▃▃▃█ beats 7/14	very hard · avg 2.3%	▇ 2.2%	sharp · at its level · noisy

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #31 of 393 * · 4.0% of its score · 1 test

Beats Claude Sonnet 5.5, GPT-6 Sol, GPT-5.6 Terra here, though they rank higher overall. Trails Claude Opus 4.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1472.7 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 345/379	avg 1268	▇▇ 4.0%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #50 of 398 · 39% of its score · 6 tests

Beats LongCat 2.0 (1.6T-A48B), Qwen3.6 Max Preview, Claude Sonnet 4.6 here, though they rank higher overall. Trails A.X K2 (688B-A33B), DeepSeek V4 Flash (Preview, Apr 2026) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tiers 1–3 v2	51.2% ✎	▂··▁▁▃▄▄█▅▃▃ beats 35/82	very hard · avg 1.2%	▇▇▇▇▇▇▇▇ 20%	sharp · at its level
FrontierMath tier 4 v2	17.1% ✎	▁▂▁▂█▅▂▂▁▁▁▁ beats 17/63	very hard · avg 0.1%	▇▇▇▇ 11%	sharp
MathArena ArXivMath January 2026 (23 problems)	58.2% ✎	▂··▄▂▄█▆▆▄▄█ beats 12/23	hard · avg 22%	▇ 3.5%	at its level
MathArena ArXivMath December 2025 (17 problems)	43.4% ✎	▃·▃▃█▃·▄·▁▁▁ beats 12/18	hard · avg 19%	▇ 2.4%	at its level
OTIS mock AIME 2024–2025	95.6% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 161/193	very hard · avg 8.5%	▇ 1.3%	sharp · noisy
FormalProofBench (200 graduate-level Lean 4 problems, agentic)	15.0% ✎	▃▅·▃··▅▃█··▃ beats 6/10	very hard · avg 3.0%	▇ 0.8%	noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #29 of 147 · 10% of its score · 5 tests

Beats Claude Sonnet 5.5, MiMo-V2.6-Pro, GPT-5.6 Terra here, though they rank higher overall. Trails Qwen3.6 Plus, Qwen3.7 Plus here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
CharXiv (reasoning questions) lab	80.3% ✎	▃·▃·▃▃▃·▃▃█▅ beats 5/11	medium · avg 40%	▇▇ 4.7%	typical
Arena vision (style controlled)	1272.9 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 118/145	avg 1068	▇ 2.7%	blunt
Arena document (style controlled)	1439.2 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 7/37	avg 1335	▇ 1.3%	blunt
Diagram-MMU diagram question answering (DQA accuracy, scientific TikZ diagrams)	84.1% ✎	▄·····▄▄·█·█ beats 4/6	medium · avg 55%	▇ 0.6%	blunt
SpaceCast-Bench predictive spatial reasoning (overall, 16 task types)	56.7% ✎	▃·▅▃▅▃··▅··█ beats 10/11	very hard · avg 32%	▇ 0.4%	blunt · at its level · noisy

web search▂▁▂▆▄█▃▃▁▁·▁ #35 of 80 * · 1.6% of its score · 1 test

Beats Gemini 3.1 Pro (Preview), Qwen3.7 Max, GLM-5.3-Flash here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena search systems (style controlled)	1191.3 ✎	▁·▁▁▂▂▂█▅▂▁▁ beats 13/30	avg 1055	▇ 1.6%	blunt

other tests8 not in a task · 21% of its score

test	result	vs the others in its run	difficulty	counts	why
ARC-AGI-2 public evaluation set (ARC Prize model_baseline, pass@2)	34.0% ✎	█··▂▅▁▃▁▃·▁▃ beats 18/24	very hard · avg 0.3%	▇▇ 4.2%	sharp · at its level · noisy
GPQA Diamond	89.4% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 168/208	hard · avg 40%	▇▇ 4.0%	sharp
Mystery Game Puzzles (Epoch)	20.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 34/74	very hard · avg 9.4%	▇ 3.6%	sharp · noisy
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	43.9% ✎	▁▁▂▁▃▄█▆▄▅▃▃ beats 36/39	very hard · avg 22%	▇ 3.0%	blunt · at its level
Chess Puzzles (Epoch)	40.0% ✎	█▁▁▁▁▂▁▄▂▃▁▁ beats 117/135	very hard · avg 5.9%	▇ 2.7%	sharp · noisy
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)	31.9% ✎	▁·▇▃▃·█▇▁·▇▃ beats 8/27	very hard · avg 4.1%	▇ 2.4%	sharp · noisy
SimpleQA Verified	66.8% ✎	▁▁▃▂▃▃▆▆█▃▃▅ beats 69/78	very hard · avg 8.0%	▇ 1.5%	at its level
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	86.5% ✎	▁▂▁▃▅█▂▃▄▃▁▁ beats 16/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $1.13/M · $0.5 in / $3 out per 1M tokens · 1M context · ai.google.dev ↗ · price sourceUSD per 1M tokens, Gemini Developer API Standard paid tier, gemini-3-flash-preview (Gemini 3 Flash Preview, which Google calls 'our legacy Flash model'; the deprecations table lists 'No shutdown date announced'): input $0.50 (text/image/video), output $3.00 (including thinking tokens). Audio input ($1.00), context caching, Batch and Flex ($0.25 in / $1.50 out) and the free tier are excluded. Currency: USD. Google Gemini API official pricing page, observed 2026-09-26.

$ how the score was worked out · Epoch AI — internal evaluatio… 51%, LMArena leaderboard dataset 17%, τ²/τ³-bench leaderboard: Sier… 7% and 9 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Epoch AI — internal evaluations	8	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 51%	CC BY 4.0
LMArena leaderboard dataset	5	▇▇▇▇▇ 17%	CC BY 4.0
τ²/τ³-bench leaderboard: Sierra's own evaluation runs	4	▇▇ 7%	MIT
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets	2	▇▇ 7%	MIT
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	2	▇▇ 6%	CC BY-SA 4.0
Google DeepMind Gemini model page and model cards (lab-reported scores) · lab-reported	1	▇ 5%	Factual vendor-reported score (not licensed content)
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 3%	CC BY 4.0
Vibe Code Bench (Vals AI, arXiv:2603.04601v3)	1	▇ 2%	CC-BY-4.0
FormalProofBench (Vals AI, arXiv:2603.26996v1)	1	▇ 1%	CC-BY-4.0
Diagram-MMU: A Multi-Modal Benchmark for Scientific Diagrams (arXiv:2608.12262v1)	1	▇ 1%	CC BY 4.0
SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)	1	▇ 0%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 72.3