Inference
EN
Google

Gemini 4 Argon

Gemini 4 Argon is a reasoning model from Google in the Gemini 4 family. 48 benchmarks count toward its index, in 5 categories.

  • Closed weights
  • Reasoning
  • Input: audio · file · image · text · video
  • Released 2026-09
RankedEnough results for an index.
IndexAll categories, out of 100.78.5±4.3Median 51.0
CoverageIndex weight that has results.80%
SpeedOutput tokens per second.28/sMedian 67/s
First tokenSeconds to the first output token.1.6 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$2.00batch $1.00Median $0.60
Output / 1MUS dollars per 1M output tokens.$10.00batch $5.00 Batch price per 1M output tokens.Median $2.55
ContextMaximum tokens in one request.1.05MMedian 262K
EloLMArena rating and rank.1530±5 #2Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
google/gemini-4-argon
Max outputMaximum output tokens per request.
944K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
Yes
Cache read / 1MUS dollars per 1M cached tokens.
$0.10
WeightsLink to the published weights.
—
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
3 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Google AI StudioMaker$2.00$10.001.05M
Google Vertex AIMaker$2.00$10.001.05M
OpenRouterRouter$2.00$10.001.05M

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.17 counted77.36 / 218
CodingCode writing and repair.13 counted76.78 / 209
ReasoningLogic problems and puzzles.10 counted79.89 / 227
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.4 counted76.57 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.None counted——
MathMath problems.4 counted75.911 / 168

Results

48 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
OTIS Mock AIME 2024-2025Math100.0%1 / 200
ARC-AGI-1 (semi-private)Reasoning98.5%1 / 126
GPQA diamond · Epoch AIKnowledge96.4% ±1.22 / 217
LiveBench ReasoningReasoning95.1%1 / 80
LiveBench Data AnalysisReasoning85.2%1 / 80
Artificial Analysis Humanity's Last ExamKnowledge57.1%4 / 222
IOICoding100.0%1 / 45
Vibe Code Bench v1.1Coding91.9% ±1.93 / 123
Code MigrationCoding68.2% ±4.32 / 77
SimpleQA VerifiedKnowledge76.0% ±1.43 / 91
Artificial Analysis SciCodeCoding61.8%5 / 114
DeepSWEAgentic77.9%3 / 59
ARC-AGI-2 (semi-private)Reasoning91.3%7 / 131
Critical Physics TasksReasoning27.1%17 / 214
LiveBench Browser UseAgentic92.0%4 / 49
Artificial Analysis AutomationBenchAgentic77.5%2 / 23
ScarfBench: Cross-Framework Application Migration in Enterprise JavaCoding75.2%1 / 11
LiveBench Long AgenticAgentic69.9%5 / 51
Mystery Game PuzzlesReasoning64.0% ±4.88 / 80
GDPval-AA normalizedAgentic56.2%14 / 133
ProofBench v1.1Math99.0% ±1.06 / 54
Chess PuzzlesReasoning59.0% ±4.910 / 90
Agent Arena task outcomeAgentic13.27 / 54
Terminal-Bench 4.0Agentic57.6% ±2.38 / 61
PostTrainBench v1.1Coding45.3%2 / 15
FrontierMath-Tiers-1-3-v2-PrivateMath86.0% ±2.113 / 90
MazeBench: visual spatial reasoning in a 3D open worldAgentic12.0%7 / 48
Artificial Analysis Omniscience AccuracyKnowledge48.3%26 / 164
OSWorld 2.0Agentic69.2%4 / 25
FrontierMath-Tier-4-v2-PrivateMath78.0% ±6.512 / 72
CWE-bench v1Agentic68.0%2 / 12
Agent Arena steerabilityAgentic13.29 / 54
ExploitGymAgentic39.5%4 / 23
ARC-AGI-3 (semi-private)Reasoning46.3%5 / 26
EBR-benchReasoning64.8% ±4.96 / 31
Artificial Analysis Long Context ReasoningReasoning79.7%45 / 217
Furniture AssemblyReasoning72.5% ±5.68 / 37
LiveBench ResearchAgentic83.1%12 / 52
cursorBench40Coding52.9%5 / 21
Multimodal Knowledge Agent BenchmarkAgentic51.3%5 / 19
CWE-bench v2Agentic65.0%3 / 11
LiveBench Coding 2.0Coding89.7%15 / 52
FrontierSWE v2Coding55.0%7 / 22
MirrorCodeCoding58.9% ±11.713 / 39
Vibe Code Bench 1-100Coding21.2% ±3.89 / 26
Bug Hunt BenchCoding40.7%9 / 25
AutomationBenchAgentic51.3%12 / 32
Agents' Last ExamAgentic39.5%9 / 21
Artificial Analysis AnalystAgentAgentic48.8%12 / 26
FrontierCode 2.0Coding26.3%14 / 27
Terminal-Bench 2.1 · Vals AIAgentic57.7% ±2.543 / 75
OSWorld 2.1 offline subset (82 tasks)Agentic74.7%6 / 10
LiveBench DesignAgentic73.7%32 / 52
Agent Arena command recoveryAgentic−21.448 / 50
ARC-AGI-1 (public eval)Reasoning99.5%1 / 116
Finance Agent (v2)Finance65.4% ±0.31 / 79
LegalBenchLegal88.3% ±0.43 / 149
MedCodeHealthcare58.8% ±2.13 / 106
HLAB-Vals-1.1Legal27.1% ±3.71 / 35
Tax Agent BenchFinance49.2% ±3.64 / 82
ARC-AGI-2 (public eval)Reasoning97.1%6 / 122
EMBFinance75.2% ±2.54 / 75
Legal Research BenchLegal54.8% ±3.55 / 90
CyberBench v1.1Other77.9% ±5.33 / 48
SRE BenchOther44.3% ±3.13 / 36
ProgramBenchCoding2.5% ±1.18 / 65
Terminal-Bench-ScienceScience44.3% ±6.06 / 44
SAGEEducation89.5% ±1.64 / 29
Public Benefits Bench v1.1Public benefits69.8%7 / 50
MedScribeHealthcare87.4% ±2.017 / 109
BioMysteryBenchScience76.3% ±0.47 / 30
MysteryMechanismScience45.5% ±3.47 / 30
Time Horizon Index: KSPOther52.5% ±8.94 / 13
FrontierMath-ErdosMath1.5% ±1.56 / 15
CUA-benchOther4.8%8 / 9
Graphwalks BFS 0K-128KReasoning99.7%—
LVBenchMultimodal91.7%—
LABBench2: An Improved Benchmark for AI Systems Performing Biology ResearchKnowledge88.8%—
GraphWalks BFS 256K–1MReasoning84.2%—
Chartography without toolsMultimodal71.6%—
Terminal-Bench-Science 0.1 with sixfold verifier timeoutAgentic57.6%—
Blueprint-Bench 2Multimodal54.4%—
RoboDojo SimAgentic34.2%—
ExtractBench Pro: long-tail document extraction challenges (LlamaIndex)Multimodal30.0%—
Axiom Bio morphology-to-molecule matchingKnowledge25.2%—
STEMBench120: six-domain scientific research progressReasoning2.4%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseVals AINo licence statedEpoch AICC BYARC Prize FoundationTerms restrict reuseLiveBenchNo licence statedLMArenaCC BY 4.0models.devMIT

History

5 releases
Index81.0 → 78.5
23 Oct 202612 Nov 2026
Output / 1M$10.00 → $10.00
23 Oct 202612 Nov 2026

Questions

How much does Gemini 4 Argon cost?
$2.00 per 1M input tokens, $10.00 per 1M output tokens.
How good is Gemini 4 Argon?
Its index is 78.5 out of 100. That is place 4 of 283 ranked models.
What is Gemini 4 Argon best at?
Knowledge: 76.5 out of 100, place 7 of 268.
How fast is Gemini 4 Argon?
28 output tokens per second, 1.6 s to the first token.
What is the context window of Gemini 4 Argon?
1.05M tokens.
Is Gemini 4 Argon open source?
No. Its weights are not published.
When was Gemini 4 Argon released?
September 2026.

Badge

Gemini 4 Argon on Inference360: place 4 of 283, index 78.5
HTML
<a href="https://inference360.ai/models/gemini-4-argon"><img src="https://inference360.ai/badges/gemini-4-argon.svg" alt="Gemini 4 Argon on Inference360: place 4 of 283, index 78.5" height="20"></a>
Markdown
[![Gemini 4 Argon on Inference360: place 4 of 283, index 78.5](https://inference360.ai/badges/gemini-4-argon.svg)](https://inference360.ai/models/gemini-4-argon)

Compare with

Claude Opus 5.578.6GPT-6 Astra78.0GPT-6.1 Sol78.0Muse Wonder Full77.3Claude Sonnet 5.577.2

Lower price

GPT-6.1 Terra77.1 · $5.00/1M

More from Google

All models
EmbeddingGemma 2UnrankedNano Banana 2.1UnrankedGemini 3.8 LiveUnrankedGemini 3.8 Live Extended ThinkingUnrankedGemini 4 Carbon82.7