Google Debuts Gemini 4 Argon AI Model With Restricted Access to Trusted Cybersecurity Experts
Developing story first seen 2 hours ago
New details on Gemini 4 Argon include Google’s performance claims, pricing and a one-million-token output limit, while access remains restricted to selected cybersecurity testers. Google says its engineers have used the model to save 300 TiB of memory across its data centres and convert code to Rust, including more than 800,000 lines in the Fuchsia OS kernel.
Google reports a 77.9 per cent score on the DeepSWE v1.1 software engineering benchmark and says Argon also leads on the Vals Index economic analysis test. Cybersecurity firm Wiz says it used the model to find a critical vulnerability in a system used by hospitals, though Google gave no further details. After testing, access is expected to begin with paid API customers and Google AI Ultra subscribers; Google has not set a timetable.
- Google has shared new benchmark, pricing and usage details for Gemini 4 Argon.
- Access is limited for now to selected cybersecurity testers.
- Wider access is planned for paid API and AI Ultra users after testing.
New here? Start with this
Google has developed a new artificial intelligence model called Gemini 4 Argon, designed to perform complex technical tasks including writing and understanding computer code. Currently, only selected cybersecurity experts and trusted researchers have access to test it, rather than the general public or most businesses.
The restricted access reflects growing attention to how powerful AI systems should be released safely. Early testers have reportedly used the model to identify critical security vulnerabilities, and Google says it has internal applications for improving its computer systems. The company plans eventually to offer the model to paying customers through its cloud services, though it has not yet announced a timeframe.
This illustrates a wider tension in the development of artificial intelligence. Tech companies must balance making powerful new tools available to drive innovation and solve problems, whilst managing risks associated with giving unrestricted access to advanced systems.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Google's decision to restrict access to trusted cybersecurity experts during the testing phase reflects responsible stewardship of a powerful technology. A model capable of identifying critical vulnerabilities poses genuine dual-use risks—the same capabilities that help defensive security experts could potentially be misused by malicious actors if widely available. Controlled access allows Google to monitor for misuse, gather high-quality feedback from experienced professionals, and refine the model's safety measures before broader release. This measured approach prioritises collective security over premature exposure of a sensitive tool.
The case against
Restricting access to a tool with significant defensive potential undermines the principle that cybersecurity expertise should be widely distributed to strengthen overall resilience. Limiting access to a curated group controlled by Google creates an artificial scarcity that concentrates power amongst the connected few, whilst potentially beneficial organisations operating with limited budgets miss crucial security advantages. Broader testing would yield more diverse feedback, surface edge cases, and democratise access to security innovation. The restriction also prioritises Google's commercial interests—charging for access rather than advancing collective security outcomes.
Full account
Google has unveiled Gemini 4 Argon, a new frontier artificial intelligence model developed to excel in specialised domains including software engineering, professional knowledge work, and cybersecurity defence. The company asserts that the model delivers superior performance across complex, multi-step tasks that require sustained reasoning. However, access remains tightly restricted during an initial testing phase, with deployment currently limited to a carefully vetted cohort of cybersecurity specialists and authorised partners.
The search giant has already integrated Gemini 4 Argon into its own internal operations at considerable scale. The model has undertaken large-scale codebase migrations across Google's infrastructure, including translating legacy C and C++ systems to Rust. In one striking example, Argon agents have processed more than 800,000 lines of code within the Fuchsia operating system's Zircon kernel. Additionally, the model's analysis of fleet-wide telemetry data has yielded substantial efficiency gains, reportedly liberating 300 terabytes of memory capacity across Google's data centre operations.
The rollout strategy reflects Google's cautious approach to frontier model deployment. Rather than releasing broadly, the company is channelling early access through its Fairwind Program, which connects the model with trusted cybersecurity defenders. Public pricing has been announced—$2 per million input tokens and $10 per million output tokens, with a 95 per cent discount on cached inputs—and the model supports substantially expanded output capabilities of up to one million tokens per response, a considerable increase from the 64,000-token limit of previous Gemini iterations. General availability is envisaged in stages, beginning with paid API users and Google AI Ultra subscribers, though no definitive timeline has been provided.
Gemini 4 Argon enters a competitive landscape in which OpenAI and other laboratories have recently advanced their own capabilities. Benchmarking comparisons position Argon favourably, notably achieving 77.9 per cent on the software engineering DeepSWE v1.1 assessment and comparable scores on knowledge-work evaluations. The model has already demonstrated practical security value; security partners using Argon have reportedly uncovered critical vulnerabilities that rival frontier models had overlooked. Google's emphasis on monitoring the model's internal reasoning processes and implementing safeguards against prompt injection and misalignment reflects broader industry concerns about safely deploying increasingly capable systems.
Where outlets differ
Source 2 supplies specific API pricing and token limits; Source 1 contains no financial or technical specifications
Source 2 names Wiz as a partner and describes vulnerability discoveries; Source 1 mentions security partnerships only in general terms
Source 1 emphasises Google's engagement with the U.S. government's voluntary pre-release process; Source 2 focuses on the phased rollout and Fairwind Program structure
Source 2 provides precise benchmark scores (77.9% on DeepSWE); Source 1 references benchmarks more broadly without numerical detail
Source 2 specifies internal efficiency metrics (300 TiB memory savings, code migration statistics); Source 1 mentions these achievements without quantification
Source 1 contextualises the launch within recent industry events (OpenAI DevDay, Kavukcuoglu's appointment); Source 2 focuses on Google's product roadmap trajectory
Coverage
- The Verge — Google announces Gemini 4 and says it’s so capable that only ‘trusted cyber defenders’ can have it right now
- Ars Technica — Google announces Gemini 4 Argon AI model, but you can’t use it yet