PickGoogle unveils Gemini 4 Argon
Google introduced Gemini 4 Argon, a new Gemini model it pitches for its hardest reasoning and problem-solving tasks.
Google has introduced Gemini 4 Argon, a new model it positions for its hardest reasoning and problem-solving tasks, with a strong focus on long, multi-step work in coding and professional domains.
What Google says Argon is good at
Google describes Argon as built for deep work on complex, real-world tasks in areas such as coding, finance, law, cybersecurity and science. The company says it outperforms other frontier models on a range of benchmarks and ranks first on the Vals Index, which measures performance on work in sectors like finance, coding, legal and tax.
Internally, Google says it has been using Argon on difficult engineering problems. In its Antigravity development environment, the company reports that teams increased agent-written code submissions by more than 35 percent while cutting code rollbacks by nearly a third. Google's launch materials also showcase multi-agent workflows in which Argon builds 3D scenes and small games by driving tools such as Blender and generating web code.
Pricing and access
Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95 percent. Developers can try it in Google AI Studio.
Why it matters for builders
The pricing is the first thing to check. For many agent workloads, a large share of input tokens is repeated context, such as system prompts, codebases or documents. A 95 percent discount on cached input can dramatically lower the real cost of long-running agents, so structure prompts to maximize cache reuse.
The internal metrics are also telling. Google highlights fewer rollbacks, not just more code, which is the right measure for coding agents: output that survives review. When you evaluate Argon against your current model, track how much of the agent's work ships without revision.
As always, benchmarks are a starting point. Test Argon on your own tasks, compare latency and total cost, and consider routing so you can mix models, using a top model for planning and cheaper models for routine steps.
Gemini 4 Argon is listed in our New AI Models category, and Google AI Studio is in APIs & Model Platforms.
Source: Google AI. This article is PickAI’s own summary and analysis based on the original reporting.
