AI
Google unveils Gemini 4 Argon, topping AI rival benchmarks but rolling out access slowly
Google's new flagship AI model outperforms competitors on most tests, though it's initially available only through a restricted program.
The New Stack
Google has announced Gemini 4 Argon, its long-awaited flagship AI model, which demonstrates strong performance across standard benchmarking tests when compared directly to offerings from OpenAI and Anthropic. The company is adopting a cautious rollout strategy, limiting initial access to participants in its Fairwind Program while gathering early feedback before wider availability.
Benchmarks show broad strengths, mixed coding results#
On most performance metrics, Gemini 4 Argon ranks competitively or ahead of rivals including OpenAI's GPT-6 Astra, Anthropic's Opus 5.5, and Fable 5.1. Google reports leading positions in 13 of 18 benchmark tests. However, the model's performance is uneven across different task categories.
Coding performance presents a complicated picture. Argon achieves 77.9% on DeepSWE v1.1—a new high-water mark—and scores 91.9% on Vibe Code Bench. Yet it trails significantly on FrontierSWE v2 and Terminal-Bench 4.0, where competitors outpace it by up to 10 points.
Knowledge and business tasks emerge as Argon's strongest domain. The model scores 51.3% on Zapier's AutomationBench, nearly nine points higher than Opus 5.5, and 84.2% on GraphWalks tests for large token windows, outpacing GPT-6 Astra by more than 12 points. Its performance on legal analysis tasks—19.6% on Harvey's Legal Agent Benchmark—is particularly notable, roughly tripling a competitor's score, though the absolute number underscores the difficulty of such specialized work.
Extended output and security focus#
A distinguishing feature is Argon's ability to generate up to one million output tokens in a single response. While most frontier models now support one million input tokens as standard, few match this output capacity. The expanded output enables the model to perform more elaborate reasoning across complex problems in a single call.
Google emphasizes Argon's cybersecurity applications. The company trained it to autonomously locate, validate, and patch software vulnerabilities. For Fairwind participants and internal teams, the model operates without restrictive cybersecurity guardrails. Google's Wiz division—acquired in March—is already deploying it through the Scan for Good initiative; the model reportedly identified a critical vulnerability in widely-used hospital software that earlier versions had overlooked. Argon scores 85.8% on Google's internal vulnerability discovery benchmark.
