‹ PublicAI IndexGoogle
The LLM benchmark aggregator.
Gemini 3.8 Flash
Strongest in Instruction following (#1 of 62), weakest in Jailbreak resistance (#193 of 266). Above par in 25 of 35 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5.1 and Gemini 4 Argon and ahead of Grok 4.5 and Muse Spark 1.1.
§ 1 · Profile
What it is good at
Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.
Human preference66−2.8#9/3541/1
Human preference66−2.8#9/3541/1
Knowledge58.7−3.8#15/1972/3
Academic knowledge61.2−2.7#5/1231/1
Document parsing58.5−4.8#15/831/1
Reasoning58−7.4#20/1955/5
Science62.8−0.7#4/1221/1
Reasoning62.8−6.2#10/1473/3
Mathematics50.2−13.6#86/1522/2
Coding52.7−13.4#63/1794/6
Agentic coding52.8−13.9#58/1734/5
Core abilities52−13.6#74/2003/4
Instruction following73.8leads#1/621/1
Language62.8−3.8#6/621/1
Data analysis26−40.1#30/321/1
General intelligence48.6−20.4#105/1802/3
Professional51.4−11#82/1942/2
Finance58.1−4.3#14/1621/1
IT operations47.5−26.3#14/331/1
Document design44.4−24#26/321/1
Biology research39.2−31.5#26/311/1
Medical56.2−10.3#32/1491/1
Cybersecurity30.4−33.5#42/471/1
Legal55−10.9#47/1731/1
Agents50.3−16#121/2983/6
Web research50.4−13.2#22/361/2
Long-horizon work38.2−35.6#26/321/1
Computer use45.1−23#47/691/2
Knowledge work59.3−11.4#51/2082/2
Safety52.3−9.1#121/3271/3
Secure code62.7−3.5#33/2681/1
Fairness56.5−13.5#54/2901/2
Toxicity avoidance53.9−5.1#129/2661/1
Harm refusal48.1−13.4#181/2901/2
Jailbreak resistance46.8−18.2#193/2661/1
§ 2 · Head to head
What it beats, and what beats it
The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for Gemini 3.8 Flash, left for the other.
§ 3 · Sources
Where the numbers come from
12 publications, 52 figures. Every one links to the page it was read from.
LMArena Text 1497
Artificial Analysis Intelligence Index 31
GDPval-AA 1445
AA-Briefcase 1203
Terminal-Bench 18.2% (mini-SWE-agent)
ARC-AGI-2 89.2%
LiveBench 57.3LiveBench · Agentic 16.5LiveBench · Reasoning 89.3LiveBench · Coding 73.3LiveBench · Browser Use 45.1LiveBench · Data Analysis 54LiveBench · Research 79.4LiveBench · Design 69.7
LiveBench 1.0 75.8LiveBench 1.0 · Reasoning 89.3LiveBench 1.0 · Coding 72.5LiveBench 1.0 · Agentic Coding 54.2LiveBench 1.0 · Mathematics 91.6LiveBench 1.0 · Data Analysis 54LiveBench 1.0 · Language 87.8LiveBench 1.0 · Instruction Following 81.4
SimpleBench 82.4%
Vals · Legal Research Bench 38.94%Vals · LegalBench 86.99%Vals · Harvey Legal Agent Benchmark 12.08%Vals · Finance Agent 61.44%Vals · TaxEval 74.45%Vals · MortgageTax 65.34%Vals · MedCode 48.13%Vals · MedScribe 84.5%Vals · BioMysteryBench 62.22%Vals · CyberBench 43.75%Vals · SRE Bench 11.07%Vals · SWE-bench Verified 80% (Mini-SWE-agent)Vals · Vibe Code Bench 78.65% (OpenHands)Vals · Code Migration 36.55%Vals · GPQA Diamond 94.44%Vals · MMLU Pro 90.22%Vals · ProofBench 48%
ParseBench · Overall 71.2%ParseBench · Tables 89.15%ParseBench · Charts 32.58%ParseBench · Content faithfulness 89.31%ParseBench · Semantic formatting 75.91%ParseBench · Visual grounding 69.03%
Enkrypt · Jailbreak risk 15.5%Enkrypt · Harmful content risk 3.9%Enkrypt · CBRN risk 36.2%Enkrypt · Toxicity risk 3.5%Enkrypt · Bias risk 67.7%Enkrypt · Insecure code risk 7.1%
Badge
[](https://publicai.io/model-index/m/gemini-3-8-flash)