Google
Google

Gemini 4 Argon

Release Date 2026-09-301M IN · 1M OUTProprietary ModelLimited accessGemini AI Pro and up

Gemini 4 Argon is Google's model for software development, long-document analysis and multi-step work involving text and visual material. Google says it expands the output available over earlier Gemini models, giving extended tasks more room for deliberation and finished deliverables. The company highlights code changes, specialist analysis and work across tools, alongside defenses against malicious instructions and actions outside the assigned scope. The combination connects extended deliberation with tool-driven changes and deliverable creation. Consult the model detail page's detailed specifications and evaluation analysis for context and output limits, reasoning settings, tool-call support, specialist document analysis and code repair.

API cost

$2IN$10OUT

Real-world usage

No.427 places0.2% share · 589B tokens/wk
767B10/12
935B10/19
1.2T10/26
1.2T11/2
911B11/9
589B11/16
SourceOpenRouterLast updated

Output Speed

Very slowper 2,000 tokens
28tok/s
First Output 5.2stotal: 76.6s
SourceOpenRouterLast updated

User Comparisons

Writing Style
Answer length
Shorter side
bottom 32%
Vocabulary diversity
Average
bottom 48%
Sentence rhythm
Average
bottom 46%
Language mixing
Average
top 49%
List usage
Very often
45% of lines
Heading usage
Occasionally
12% of lines
Reasoning share
Minimal
2% of total
Response Speed
instability ±29%
First response 4.5sTotal 26.9s
49 tok/s
15.8ms per token
SourceAIB Inc.Last updated

Expert Performance Evaluation

50.4Reasoning effort: High152077.752.6

By effortMedium 47.6High 50.4

More about AA Intelligence Index v5

39.5%Reasoning effort: MaxSelf-reported by developer$13,71817.1%58.582.8

More about Agents' Last Exam

79.7%85.2151482.4%53.6%65.4%49.2%54.8%27.1%58.8%69.267.7%28.4%19.6%

More about Long Context Reasoning

149276.0%12.1%86.844.3%87.9%

More about Arena Text Factuality

57.1%15522100168261.8%85.747.6%77.9%15472.5%91.9%56.5%55.0%58.9%78.7%40.4%

More about Terminal-Bench 4.0

1521154215201526123687.4%1513

More about Arena Writing, Literature & Language

57.1%27.1%1535100.0%86.0%78.0%99.0%1540154295.11.5%98.5%91.3%91.7%59.0%64.0%97.6%63.7

More about Humanity's Last Exam

72.5%Reasoning effort: High54.4%71.6%

More about Furniture Assembly

Specialized fields

14871522154415491534152764.8%78.6%73.4%54.2%71.5%72.2%84.0%89.9%72.4%95.4%96.3%94.5%137191653675.6%

More about Arena Korean

OddBench

66.0%Reasoning effort: Low98.6%44645039.3168610831121

By effortLow 66.0%High 65.0%

More about BullshitBench v2

See each evaluation’s detail page for its source and last updateData corrections & removal requests