AIEvals

Models

Gemini 4 Argon

Google DeepMind’s Gemini 4 model for long-running coding, knowledge work and multimodal reasoning, initially available to trusted testers and cyber defenders.

GoogleReleased 30 September 2026Context 1M tokensOutput 1M tokensPrivate weights$2.00 in / $10.00 out per Mtok

Family Gemini 447 benchmarks with a published result4 independent first places among the 46 models this site listsRead 17 October 2026

Where it stands

Rank compares tracked models with a score. Published shows the source’s rank and field size, when available.

BenchmarkScoreRankPublishedLeaderBehind
Vals IndexVals AI68.90%1of 331of 41Gemini 4 Argon—
Artificial Analysis Intelligence IndexArtificial Analysis533of 42no list readClaude Opus 5.55 pts
Terminal-Bench 4.0Artificial Analysis57.07%4of 40no list readClaude Sonnet 5.56.57 pts
Vibe Code Bench v1.1Vals AI91.91%2of 382of 106Claude Sonnet 5.50.48 pts
ProgramBenchVals AI2.50%5of 295of 54Claude Opus 5.516.00 pts
Vals RSI IndexVals AI30.55%4of 204of 23Claude Opus 5.56.76 pts
CUA-benchVals AI4.83%7of 87of 8GPT-6 Astra14.34 pts
AutomationBench-AAArtificial Analysis77.51%1of 40no list readGemini 4 Argon—
SciCodeArtificial Analysis61.81%3of 40no list readClaude Opus 5.55.09 pts
GDP.pdfArtificial Analysis21.80%17of 40no list readGPT-6.1 Sol9.20 pts
CritPtArtificial Analysis27.14%11of 42no list readGPT-5.6 Sol5.15 pts
AA-OmniscienceArtificial Analysis42.355of 42no list readClaude Opus 5.54.07
AA-LCR v1.1Artificial Analysis79.67%30of 42no list readKimi K39.00 pts
ProofBench v1.1Vals AI99.00%4of 345of 46Claude Sonnet 5.51.00 pts
LegalBenchVals AI88.30%3of 273of 149Claude Fable 50.26 pts
Harvey's Legal Agent BenchmarkVals AI19.58%3of 375of 73Muse Spark 1.25.84 pts
MysteryMechanismVals AI45.49%6of 216of 21GPT-6 Astra7.66 pts
BioMysteryBenchVals AI76.30%6of 196of 21Claude Sonnet 5.54.81 pts
Code MigrationVals AI68.17%2of 372of 71Claude Sonnet 5.51.66 pts
CyberBench v1.1Vals AI77.86%2of 312of 43GPT-6 Sol0.12 pts
EMBVals AI75.24%4of 374of 69Claude Fable 5.11.43 pts
Finance Agent (v2)Vals AI65.40%1of 371of 73Gemini 4 Argon—
IOIVals AI100.00%1of 311of 38Gemini 4 Argon—
Legal Research BenchVals AI54.81%4of 374of 72Muse Spark 1.3 Max0.48 pts
MedCodeVals AI58.80%3of 353of 104Claude Opus 54.77 pts
MedScribeVals AI87.43%13of 3615of 106Claude Opus 5.54.00 pts
Public Benefits Bench v1.1Vals AI69.76%5of 295of 45Claude Opus 57.17 pts
SAGEVals AI53.65%4of 325of 90Claude Opus 4.72.45 pts
SRE BenchVals AI44.27%3of 153of 16GPT-6 Astra12.60 pts
Tax Agent BenchVals AI76.23%2of 372of 64Claude Fable 5.11.41 pts
GDPval-AA v2.1Artificial Analysis1,611.3314of 42no list readClaude Opus 5.5234.84
AA-Briefcase v1.1Artificial Analysis1,493.8418of 41no list readClaude Opus 5.5328.01
Humanity's Last ExamArtificial Analysis57.09%3of 42no list readClaude Opus 5.54.26 pts

Nearby on Vals Index

The two models above and below on the Vals Index (Vals AI), on every column where this model has a published number.

ModelVals Index%Artificial Analysis Intelligence IndexindexTerminal-Bench 4.0%Vibe Code Bench v1.1%ProgramBench%
Gemini 4 ArgonGoogle68.90%5357.07%91.91%2.50%
Claude Sonnet 5.5Anthropic67.04%5663.64%92.39%–
Claude Opus 5.5Anthropic66.97%5859.60%90.29%18.50%
Bold marks a figure better than this model’s on the same column, whichever model it belongs to. Ranks cover the models shown from the stated source.

Cost and speed, as published

Additional measurements from Artificial Analysis’ own row. Read 1 October 2026.

As publishedCost per taskOutput speed
Gemini 4 Argon (High)$1.99—

First answer token: —; 500-token response: —. AA publishes no API speed or latency measurements for this restricted-access model in the checked snapshot. Provider measurements

Results by benchmark

Independent and vendor results retain their own harnesses and effort settings.

BenchmarkIndependent runAs reported by a publisher
Vals RSI IndexAutonomous research30.55%Vals AInot published
CUA-benchComputer use4.83%Vals AInot published
LABBench 2Scientific reasoningno published result
88.8%LABBench 2Google DeepMind · October 2026Rank 1 of 4 in this publisher table
RiemannBenchMathematicsno published result
76.0%RiemannBenchGoogle DeepMind · October 2026Rank 1 of 4 in this publisher table
GraphWalks (up to 128k)Long contextno published result
99.7%GraphWalks (Up to 128k, BFS F1)Google DeepMind · October 2026Rank 1 of 4 in this publisher table
GraphWalks (256k to 1M)Long contextno published result
84.2%GraphWalks (256k to 1M, BFS F1)Google DeepMind · October 2026Rank 1 of 4 in this publisher table
LVBenchVideo understandingno published result
91.7%LVBench (no tools)Google DeepMind · October 2026Rank 1 of 4 in this publisher table
CWE-bench v1Securityno published result
68.0%CWE-bench v1Google DeepMind · October 2026Rank 1 of 4 in this publisher table
AutomationBench-AAWorkflow automation77.51%Artificial Analysisnot published
SciCodeScientific coding61.81%Artificial Analysisnot published
GDP.pdfDocument reasoning21.80%Artificial Analysisnot published
CritPtScientific reasoning27.14%Artificial Analysisnot published
AA-OmniscienceFactual knowledge42.35Artificial Analysisnot published
AA-LCR v1.1Long-context reasoning79.67%Artificial Analysisnot published
Terminal-Bench 4.0Agentic coding57.07%Artificial Analysis
57.58%Vals AI
57.4%Terminal-Bench 4.0Google DeepMind · October 2026Rank 4 of 4 in this publisher table
Vibe Code Bench v1.1Product coding91.91%Vals AI
91.9%Vibe Code BenchGoogle DeepMind · October 2026Rank 1 of 4 in this publisher table
Terminal-Bench Science 0.1Agentic science
44.29%Vals AI
57.6%Terminal-Bench Science 0.1Google DeepMind · October 2026Rank 3 of 4 in this publisher table
ProgramBenchAgentic coding2.50%Vals AInot published
ProofBench v1.1Mathematics99.00%Vals AInot published
LegalBenchProfessional knowledge88.30%Vals AInot published
Harvey's Legal Agent BenchmarkProfessional knowledge19.58%Vals AI
19.6%Harvey's Legal Agent BenchmarkGoogle DeepMind · October 2026Rank 1 of 4 in this publisher table
MysteryMechanismScientific reasoning45.49%Vals AInot published
Vals IndexComposite68.90%Vals AI
68.9%Vals IndexGoogle DeepMind · October 2026Rank 1 of 4 in this publisher table
BioMysteryBenchAgentic science76.30%Vals AInot published
Code MigrationAgentic coding68.17%Vals AInot published
CyberBench v1.1Security77.86%Vals AInot published
EMBFinancial modelling75.24%Vals AInot published
Finance Agent (v2)Financial analysis65.40%Vals AI
65.4%Vals Finance Agent v2Google DeepMind · October 2026Rank 1 of 4 in this publisher table
IOICompetitive programming100.00%Vals AInot published
Legal Research BenchProfessional knowledge54.81%Vals AInot published
MedCodeHealthcare administration58.80%Vals AInot published
MedScribeHealthcare administration87.43%Vals AInot published
Public Benefits Bench v1.1Public services69.76%Vals AInot published
SAGEEducation53.65%Vals AInot published
SRE BenchSecurity44.27%Vals AInot published
Tax Agent BenchProfessional knowledge76.23%Vals AInot published
OSWorld 2.0Computer useno published result
69.2%OSWorld 2.0 (offline set, partial reward)Google DeepMind · October 2026Rank 2 of 2 in this publisher table
AutomationBenchBusiness workflowsno published result
51.3%AutomationBenchGoogle DeepMind · October 2026Rank 1 of 4 in this publisher table
DeepSWE v1.1Software engineeringno published result
77.9%DeepSWE v1.1Google DeepMind · October 2026Rank 1 of 4 in this publisher table
GDPval-AA v2.1Knowledge work1,611.33Artificial Analysisnot published
AA-Briefcase v1.1Knowledge work1,493.84Artificial Analysisnot published
ChartographyVisual chart recognitionno published result
71.6%Chartography (no tools)Anthropic · 7 October 2026Rank 1 of 10 in this publisher table
71.6%Chartography (no tools)Google DeepMind · October 2026Rank 1 of 4 in this publisher table
Humanity's Last ExamMultidisciplinary reasoning57.09%Artificial Analysisnot published
Agents' Last ExamComputer useno published result
39.5%Agent's Last Exam (binary pass rate)Google DeepMind · October 2026Rank 1 of 3 in this publisher table
Artificial Analysis Intelligence IndexComposite53Artificial Analysisnot published
FrontierSWEAgentic researchno published result
55.0%FrontierSWE v2Google DeepMind · October 2026Rank 4 of 4 in this publisher table
PostTrainBenchAI researchno published result
45.3%PostTrainBench v1.1Google DeepMind · October 2026Rank 2 of 4 in this publisher table
Independent column: the suite named under each figure, one harness per suite. Publisher column: each lab’s own table, reproduced on the benchmark’s page. Publisher ranks compare reported scores in the same table row only; ties share a rank and unreported cells are excluded.

Source notes

From the linked sources. Read 1 October 2026.

Restricted access. Announced on 30 September 2026 with phased access for trusted testers and cyber defenders through the Fairwind Program. Broad access for developers, enterprises and consumers has not yet launched in the checked announcement.

Introductory pricing. Google announces $2 per million input tokens and $10 per million output tokens for launch, with cached input discounted by 95%. After the introductory period, prices become $4 / $20; the announcement does not specify when that period ends.

Input context. Vals AI and Artificial Analysis record a 1M-token input context. Google’s checked announcement and evaluation do not state the public API input-context limit.

Evaluator configuration. Vals AI records a 262,144-token maximum output, temperature 1 and high reasoning effort, with $4 / $20 token pricing. These are its evaluation settings; Google’s announcement specifies a 1M-token output limit and introductory $2 / $10 pricing. AA evaluates High effort at the introductory prices. Independent results retain each evaluator’s harness, version and snapshot date.

Evaluation settings. The October 2026 report uses the Gemini API at its highest thinking setting, generally pass@1 with no majority voting or parallel test-time compute. Its 19 results remain attributed to Google DeepMind’s published table, including figures it quotes from external leaderboards.

Sources

Read on 17 October 2026.