GoogleReasoning model

Gemini 3.8 Flash

Reasoning effort

Gemini 3.8 Flash is a Google proprietary reasoning model, released 2 Sept 2026. Its best configuration (medium reasoning effort) ranks #48 of 751 on the BenchLeader Index at 64.7 ±4.6, in the upper half. The ± is the point: 118 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in agents & tools (75) and lowest in coding (58). At $1.50 per million tokens blended it is pricier than most ranked models. Output speed of 75 tokens per second puts it slower than most, with a first token in 2.5 s. It has been measured at 4 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 2 Sept 2026.

Blended price: $1.50/M; $0.750 in · $3.75 out
Output speed: 75 tok/s; OpenRouter traffic, 7-day median; not yet measured by Artificial Analysis
First answer: 2.47 s
Context: 1.0M
Full answer: –
Cost per run: $0.931; one full Intelligence Index run
Released: 2 Sept 2026

How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.

Overall index65
Reasoning63
Coding58
Agents & tools75
Knowledge75
Multimodal68
Long context68
Composite75

Versions

Google has shipped 13 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.

Model	Released	Index	Rank
Gemini 3.8 Flashmediumthis page	2 Sept 2026	64.7	#48
Gemini 3.7 Flashmedium	13 Aug 2026	63.5	#63
Gemini 3.6 Flashhigh	21 Jul 2026	60.5	#102
Gemini 3.5 Flashmedium	19 May 2026	63.1	#68
Gemini 3 Flashthinking	17 Dec 2025	59.3	#122
Gemini 2.5 Flash 09 2025	25 Sept 2025	51.9	#314
Gemini 2.5 Flash	17 Apr 2025	52.8	#283
Gemini 2.0 Flashthinking	5 Feb 2025	45.0	#496
Gemini 1.5 Flash 002	24 Sept 2024	40.7	#623
Gemini 1.5 Flash 001	23 May 2024	36.4	#717
Gemini 1.5 Flash 0514	14 May 2024	–	–
Gemini 1.5 Flash	–	–	–
Gemini Flash 2.0	–	–	–

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Not stated” collects results from publishers that did not say which setting they used; for a reasoning model that is usually its thinking mode, but we do not assume it. Pick a setting here or at the top of the page to see its price, speed and category scores.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
low	61.3	#94	75 tok/s	2.47 s	$0.0014	Agents & tools 62 · Coding 61 · Composite 68 · Knowledge 71 · Long context 66 · Multimodal 68 · Reasoning 57
mediumbest	64.7	#48	75 tok/s	2.47 s	$0.0014	Agents & tools 75 · Coding 58 · Composite 75 · Knowledge 75 · Long context 68 · Multimodal 68 · Reasoning 63
high	63.7	#61	249 tok/s	19 s	$0.0014	Agents & tools 55 · Coding 62 · Composite 64 · Human preference 69 · Knowledge 69 · Long context 66 · Maths 63 · Multimodal 68 · Reasoning 69
not stated	58.6	#142	75 tok/s	2.47 s	$0.0014	Agents & tools 61 · Knowledge 60 · Maths 55 · Reasoning 76

How its index has moved

8 Sept 2026 to 4 Oct 2026

medium
high
low
not stated

The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.

Benchmark results

One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Reasoning

Benchmark	low	medium	high	not stated	Source
GPQA Diamond	–	–	95.4%#3	–	Epoch AI Benchmarking Hub
Humanity's Last Exam	–	–	–	44.5%#4	Scale AI / CAIS
LMArena Hard Prompts	–	–	1516#11	–	LMArena
LiveBench Reasoningnot in index	–	–	89.3%#17	–	LiveBench
GPQA Diamond (AA)not in index	92.0%#41	93.5%#14	95.3%#3	–	Artificial Analysis
Humanity's Last Exam (AA)not in index	37.1%#97	42.1%#66	47.8%#34	–	Artificial Analysis
GPQA Diamond (Vals)not in index	–	–	94.4%#4	–	Vals AI
ARC-AGI-1	90.5%#65	97.5%#10	98.5%#1	–	ARC Prize
ARC-AGI-2	77.5%#40	82.9%#36	89.2%#17	–	ARC Prize
ARC-AGI-3	15.1%#16	24.2%#14	35.0%#12	–	ARC Prize
CritPt	4.0%#149	12.3%#94	18.3%#62	–	Artificial Analysis
MysteryMechanismnot in index	–	–	36.5%#8	–	Vals AI
Chess Puzzlesnot in index	–	–	61.0%#4	–	Epoch AI Benchmarking Hub
Mystery Game Puzzlesnot in index	–	–	47.0%#12	–	Epoch AI Benchmarking Hub
LMCAnot in index	51.2%#47	52.9%#37	–	51.5%#44	Epoch AI Benchmarking Hub
DTBenchnot in index	94.9%#33	94.9%#33	–	95.7%#27	Epoch AI Benchmarking Hub

Coding

Benchmark	low	medium	high	not stated	Source
SciCode	54.3%#48	54.4%#47	56.6%#23	–	SciCode
FrontierCode	–	41.2%#21	–	–	Cognition
LMArena Coding	–	–	1530#21	–	LMArena
LMArena WebDev	–	–	1583#28	–	LMArena
LiveBench Codingnot in index	–	–	72.5%#53	–	LiveBench
SciCode (AA)not in index	55.0%#50	55.1%#48	56.6%#32	–	Artificial Analysis
LiveCodeBench	–	–	89.5%#3	–	Vals AI
IOI	–	–	56.9%#16	–	Vals AI
SWE-bench (Vals)not in index	–	–	80.0%#24	–	Vals AI
Code Migrationnot in index	–	–	36.5%#30	–	Vals AI
ProgramBenchnot in index	–	–	1.0%#11	–	Vals AI
Vibe Code Bench 1-100not in index	–	–	18.8%#9	–	Vals AI
Vibe Code Bench v1.1not in index	–	–	78.7%#23	–	Vals AI
SWE Atlas: Codebase QnAnot in index	–	–	–	47.0%#6	Scale AI SEAL
SWE Atlas: Refactoringnot in index	–	–	–	44.8%#7	Scale AI SEAL
SWE Atlas: Test Writingnot in index	–	–	–	53.7%#4	Scale AI SEAL
Surface Evolver Benchnot in index	–	–	76.9%#9	–	Epoch AI Benchmarking Hub
DeepSWEnot in index	–	71.0%#10	73.8%#2	–	Epoch AI Benchmarking Hub
CursorBenchnot in index	–	37.3%#28	39.6%#24	–	Cursor
ALE-Benchnot in index	–	–	1269.7#30	–	Epoch AI Benchmarking Hub
FrontierSWEnot in index	–	–	19.6%#13	–	Epoch AI Benchmarking Hub
Terminal-Bench 4.0 (AA)not in index	10.1%#76	19.7%#54	19.7%#54	–	Artificial Analysis
Terminal-Bench 2.1 (AA)not in index	83.2%#36	83.9%#34	87.6%#18	–	Artificial Analysis

Agents & tools

Benchmark	low	medium	high	not stated	Source
Terminal-Bench	–	–	19.1%#65	–	Terminal-Bench
APEX-Agents	–	–	–	64.3%#8	Mercor
LMArena Agent	–	–	3#20	–	LMArena
LiveBench Agentic Codingnot in index	–	–	54.2%#30	–	LiveBench
Terminal-Bench 2.1 (Vals)	–	–	81.3%#8	–	Vals AI
HiL-Bench	–	–	–	41.5%#6	Scale AI SEAL
GDPval (AA)not in index	39.7%#90	45.4%#62	45.6%#61	–	Artificial Analysis
τ²-Bench Banking (AA)	33.2%#55	45.8%#12	45.0%#15	–	Artificial Analysis
ITBench SRE (AA)not in index	–	–	52.5%#3	–	Artificial Analysis
CUA-benchnot in index	–	–	4.2%#8	–	Vals AI
CyberBenchnot in index	–	–	43.8%#36	–	Vals AI
Finance Agent v2not in index	–	–	61.4%#2	–	Vals AI
Harvey's Legal Agent Benchmarknot in index	–	–	10.0%#15	–	Vals AI
Legal Research Benchnot in index	–	–	38.9%#26	–	Vals AI
SkillsBenchnot in index	–	–	58.0%#12	–	Vals AI
Tax Agent Benchnot in index	–	–	66.8%#14	–	Vals AI
Terminal-Bench 4.0 (Vals)not in index	–	–	19.2%#22	–	Vals AI
Terminal-Bench Sciencenot in index	–	–	8.6%#14	–	Vals AI
Time Horizon Index: KSPnot in index	–	–	18.8%#4	–	Vals AI
Vending-Bench 2not in index	–	–	–	5093.8#26	Andon Labs
AutomationBenchnot in index	–	–	59.9%#21	–	Artificial Analysis
GDP.pdfnot in index	–	–	21.0%#25	–	Artificial Analysis
AA-Briefcasenot in index	–	–	1203#39	–	Artificial Analysis
FACTS Searchnot in index	–	–	–	76.3%#14	Google

Maths

Benchmark	low	medium	high	not stated	Source
FrontierMath Tiers 1–3	–	–	68.4%#26	–	Epoch AI Benchmarking Hub
FrontierMath Tier 4	–	–	21.9%#45	–	Epoch AI Benchmarking Hub
OTIS Mock AIME	–	–	98.9%#16	–	Epoch AI Benchmarking Hub
ProofBench	–	–	–	48.0%#27	Vals AI
LiveBench Mathematicsnot in index	–	–	91.6%#26	–	LiveBench
LMArena Maths	–	–	1518#5	–	LMArena

Knowledge

Benchmark	low	medium	high	not stated	Source
SimpleQA Verified	–	–	69.7%#7	–	Epoch AI Benchmarking Hub
LiveBench Data Analysisnot in index	–	–	54.0%#62	–	LiveBench
AA-Omniscience	21.4#57	28.6#36	29.6#33	–	Artificial Analysis
MMLU-Pro	–	–	90.2%#5	–	Vals AI
LegalBench	–	–	87.0%#7	–	Vals AI
TaxEval	–	–	74.5%#35	–	Vals AI
PRBench Finance	–	–	–	49.5%#10	Scale AI SEAL
PRBench Legal	–	–	–	48.9%#12	Scale AI SEAL
BioMysteryBenchnot in index	–	–	62.2%#16	–	Vals AI
Excel Modeling Benchmarknot in index	–	–	72.2%#8	–	Vals AI
MedCodenot in index	–	–	48.1%#29	–	Vals AI
MedScribenot in index	–	–	84.5%#32	–	Vals AI
Public Benefits Benchnot in index	–	–	65.3%#18	–	Vals AI
GDP.pdfnot in index	–	23.4%#17	23.2%#18	–	Epoch AI Benchmarking Hub
AA-Omniscience: accuracynot in index	52.2%#51	53.0%#49	54.6%#39	–	Artificial Analysis
AA-Omniscience: non-hallucinationnot in index	35.4%#173	48.1%#116	44.8%#130	–	Artificial Analysis

Instruction following

Benchmark	low	medium	high	not stated	Source
LiveBench Languagenot in index	–	–	87.8%#7	–	LiveBench
LiveBench Instruction Followingnot in index	–	–	81.4%#1	–	LiveBench
LMArena Instruction Followingnot in index	–	–	1486#15	–	LMArena

Human preference

Benchmark	low	medium	high	not stated	Source
LMArena Text	–	–	1495#8	–	LMArena
LMArena Creative Writingnot in index	–	–	1485#7	–	LMArena
LMArena Multi-turnnot in index	–	–	1501#8	–	LMArena
LMArena Longer Queriesnot in index	–	–	1509#8	–	LMArena

Multimodal

Benchmark	low	medium	high	not stated	Source
LMArena Vision	–	–	1316#5	–	LMArena
MMMU-Pro	84.5%#19	84.2%#21	85.6%#9	–	Artificial Analysis
MMMU-Pro (Vals)not in index	–	–	89.1%#4	–	Vals AI
MortgageTaxnot in index	–	–	65.3%#43	–	Vals AI
SAGEnot in index	–	–	35.1%#66	–	Vals AI
AudioMCnot in index	–	–	60.4%#1	–	Scale AI SEAL
Blueprint-Bench 2not in index	–	–	–	38.6%#3	Epoch AI Benchmarking Hub

Long context

Benchmark	low	medium	high	not stated	Source	Trend
AA-LCR	80.7%#58	84.0%#11	81.3%#50	–	Artificial Analysis
MLCRnot in index	–	–	21.7%#16	–	Artificial Analysis

Composite

Benchmark	low	medium	high	not stated	Source
Epoch Capabilities Indexnot in index	–	–	–	156.9#15	Epoch AI Benchmarking Hub
LiveBench	–	–	75.8%#33	–	LiveBench
AA Intelligence Index	33.5#95	39.8#58	40.9#52	–	Artificial Analysis
Vals Indexnot in index	–	–	54.8#15	–	Vals AI

Where to run it

Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.

Provider	Speed	First token	Input $/M	Output $/M	Context	Quantisation
Google AI Studio	176 tok/s	1.68 s	$0.375	$1.88	1.0M	–
Google AI Studio	172 tok/s	1.59 s	$1.35	$6.75	1.0M	–
Google AI Studio	150 tok/s	1.50 s	$0.750	$3.75	1.0M	–
Google Vertex	95 tok/s	3.26 s	$1.35	$6.75	1.0M	–
Google Vertex	92 tok/s	2.19 s	$0.750	$3.75	1.0M	–
Google Vertex	9 tok/s	11 s	$0.375	$1.88	1.0M	–

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.075 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0014	$0.0012	6.5 s
Summarise a 30-page report	12,000 / 600	$0.011	$0.0052	10.5 s
Code edit	6,000 / 1,500	$0.010	$0.0071	22.5 s
Agentic coding session	60,000 / 4,000	$0.060	$0.030	55.8 s
Structured extraction	2,000 / 200	$0.0023	$0.0012	5.1 s