Model
Model explorer

Gemini 4 Argon

CLOSED
Google · Gemini 4 family · released Sep 30, 2026

Google DeepMind's Sep 30, 2026 frontier model, announced by Koray Kavukcuoglu and initially rolling out only to trusted cyber defenders via the Fairwind Program, with paid API and AI Ultra access promised later. Google reports DeepSWE v1.1 77.9%, AutomationBench 51.3% (#1), LVBench 91.7% and a tie for first on CWE-bench v1 at 68.0%, with an output token limit raised to 1M. Announced introductory pricing is $2/$10 per Mtok, rising to $4/$20 after the introductory period. It trails competitors in Google's own table on FrontierSWE v2, Terminal-Bench 4.0, PostTrainBench, Terminal-Bench Science and OSWorld 2.0. Corpus choices: ctxK is filed as 1000 because Google publishes only the 1M-token output limit (the input window is at least that); price is null until the announced $2/$10 introductory price is self-serve, following the Gemini 3.8 Flash Cyber precedent; status released with apiAvailable false for the same reason.

ReasoningCodingVisionFunction callingTool useAgentic
3254.0
Elo · rank #2
Parameters
Undisclosed
Active params
—
Context
1M tokens
Architecture
Proprietary model; architecture and parameter count undisclosed; 1M-token output limit; gated to trusted cyber defenders at launch
License
Proprietary (Fairwind Program trusted-defender access)
Languages
—
API price (in/out)
No hosted API
Modalities
text · vision · video
Benchmark results
Bar shows position within the tracked field; marker = field best
AutomationBenchAgents51.3%#4
best: DeepSeek-V4.1-Flash · 54.8%
ChartographyVision71.6%#1
best: this model · 71.6%
DeepSWECoding77.9%#1
best: this model · 77.9%
FrontierSWE v2Coding55.0%#3
best: Claude Opus 5.5 · 62.3%
best: this model · 19.6%
LVBenchVision91.7%#1
best: this model · 91.7%
OSWorld 2.0Agents69.2%#4
best: Claude Opus 5.5 · 81.8%
Terminal-Bench 4.0Agents57.4%#5
best: Claude Sonnet 5.5 · 70.6%
Vals Finance AgentAgents65.4%#1
best: this model · 65.4%
Run it locally
Closed weights — available via API only. No local deployment.
Input / M tok
—
Output / M tok
—
Gemini 4 family
Elo progression across releases
API price — · each benchmark row carries its own source badge (see methodology)