Google’s new frontier AI model is designed for long-horizon software engineering, enterprise knowledge work and cybersecurity defence, with early access beginning through its Fairwind Program.
5 October 2026 — Google has introduced Gemini 4 Argon, its latest frontier artificial intelligence model, targeting complex professional workflows across software engineering, finance, legal work and cybersecurity.
The model significantly expands Gemini’s capacity for long-running tasks, increasing its maximum output from 64,000 tokens to an industry-leading 1 million tokens, according to Google. The larger output window is intended to allow the model to sustain deeper reasoning and execute multi-stage workflows without repeatedly breaking tasks into smaller interactions.
Gemini 4 Argon is initially being made available to selected cybersecurity organisations through Google DeepMind’s Fairwind Program, with Google taking a phased approach before making the model more broadly available to developers, enterprises and consumers.
Gemini 4 Argon targets long-horizon enterprise work
Google is positioning Argon around tasks where AI systems need to maintain context, reason across many steps and produce substantial outputs rather than simply answer individual prompts.
That includes large software migrations, financial research, legal drafting, cybersecurity remediation and other knowledge-intensive enterprise activities.
On DeepSWE v1.1, a benchmark focused on real-world long-horizon software engineering, Google reports that Gemini 4 Argon achieved a score of 77.9%, setting a new state-of-the-art result on the benchmark.
Google says thousands of its employees are already using Argon internally across specialised coding and research workflows, providing an early indication of how the company expects the model to be used in production environments.
Google puts Argon to work inside its own infrastructure
Some of the most notable examples come from Google’s own engineering operations.
The company says teams have used Argon agents to analyse fleet-wide profiling data and identify memory optimisations across its data centres. Once deployed, those changes are expected to free more than 300 TiB of memory, with Google estimating potential total savings between 500 TiB and 1 PiB.
Argon is also being applied to large-scale codebase migrations from C and C++ into Rust.
Google says the model has worked on projects ranging from tens of thousands of lines of code in core libraries to more than 800,000 lines within the Zircon kernel used by Fuchsia OS. The company says these migrations remain subject to automated testing, emulation and human review before deployment.
In another example, Argon agents reworked 32,000 lines of SIMD code in the Rust version of Google’s libgav1 video decoder. Google reports that the resulting implementation ran 2.7 times faster than the previous Rust port while producing identical video output.
Cybersecurity becomes a major focus
Cybersecurity is another central part of the release.
Google says Gemini 4 Argon has been trained to autonomously discover, validate and patch software vulnerabilities.
On CWE-bench v1, which evaluates vulnerability remediation, Argon scored 68%, tying for the highest reported score on the benchmark alongside other leading frontier models.
Through the Fairwind Program, approved cybersecurity teams can access versions of Argon with advanced defensive capabilities.
Google DeepMind says the programme currently works with more than 650 partners globally, prioritising governments, critical infrastructure operators and major technology platforms. Access is controlled through organisational vetting, authentication requirements and restrictions around permitted use.
The company says malicious activities such as malware creation are prohibited, while authorised activities including threat simulation, malware analysis and vulnerability research may be permitted for defensive purposes.
From short prompts to extended AI workflows
The move to a 1 million-token output limit reflects a broader shift in frontier AI development.
Earlier generative AI systems were primarily designed around individual prompts and relatively short responses. Newer models are increasingly being built for work that may involve analysing large datasets, running tools, generating substantial codebases and carrying out tasks over long periods.
For businesses, that could make AI more useful in areas that previously required repeated human intervention.
An engineering team could potentially assign an AI system a larger migration or optimisation task rather than asking for isolated code snippets. Similarly, financial and legal professionals could use models for deeper research and document-intensive workflows.
However, larger autonomous workflows also increase the importance of verification, governance and human oversight, particularly when models are working with critical infrastructure or sensitive business systems.
Google takes a phased approach to availability
Google is not releasing Gemini 4 Argon universally at launch.
Instead, the company says it is gathering feedback from selected testers and working through safety evaluations before widening access.
Google is also participating in the US government’s voluntary process for pre-release model access as part of its approach to frontier AI deployment.
The company has announced an introductory API price of $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95% from the standard input price.
Broader availability is expected to extend eventually to developers, enterprise customers and consumers.
Enterprise AI moves toward longer autonomous tasks
Gemini 4 Argon arrives as competition among frontier AI developers increasingly shifts away from chatbot performance alone.
The focus is moving toward how well models can complete real work: maintaining context, using tools, writing and modifying software, analysing enterprise data and operating reliably across longer sequences of decisions.
Google’s internal use cases show where that competition may be heading.
The value of frontier models may increasingly be measured not by how well they answer a single question, but by how much complex work they can complete before requiring intervention.
For companies considering the next generation of enterprise AI, Gemini 4 Argon therefore represents another step toward systems designed to work alongside professional teams over extended workflows rather than simply act as conversational assistants.
About Google DeepMind
Google DeepMind is Google’s artificial intelligence research and development organisation, working across frontier models, AI science, robotics, cybersecurity and other advanced AI applications.
Its Gemini model family powers AI capabilities across Google products and enterprise services.
