Google parent Alphabet has launched Gemini 4 Argon, a new frontier model that the company calls its most powerful yet. Built to sustain deep reasoning across long, multi-step workflows, Argon is positioned as a workhorse for real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. But unlike previous Gemini releases, Argon is not available to the public. It is rolling out first to a select group of trusted cyber defenders through Google's Fairwind Program, with wider access promised later.
The release comes nearly a year after Gemini 3 put Google back at the forefront of the AI race, and after months of delays that saw the company skip its promised Gemini 3.5 Pro. With Argon, Google is making a deliberate bet on long-horizon reasoning and specialized capability rather than another incremental chatbot upgrade. The model's headline change is an industry-leading 1 million token output limit, up from 64,000 tokens in previous Gemini models. That expanded headroom allows Argon to think deeply and generate hundreds of thousands of tokens in a single trajectory, which Google says adds a new level of depth to solving tough problems in one go.
A model for deep, long-horizon work
Google has backed its claims with a broad set of benchmark results. On DeepSWE v1.1, which measures performance on real-world software engineering tasks, Argon scores 77.9%, ahead of OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models. On the Vals Index, which weights contributions to U.S. GDP across finance, coding, legal, and tax work, Argon leads with 68.9%, just ahead of Claude Opus 5.5 at 67.0%. It also ranks first on Zapier's AutomationBench with a score of 51.3%, and on LVBench, a long-video understanding benchmark, Argon hits 91.7%.
Those numbers tell a specific story. Argon is not necessarily the best model at every task. Independent analysts note that it trails Claude Opus 5.5 on some ML engineering workloads and shows average performance in creative writing and everyday coding. But on long, multi-step, tool-heavy work, Argon appears to hold a clear edge. That is exactly the kind of work that matters for enterprise adoption and for autonomous agents.
Cyber defense is the priority
Cybersecurity is where Google is focusing the launch. The company says Argon was specifically trained for defensive cyber work and can autonomously find, validate, and patch critical software vulnerabilities. On CWE-bench v1, a benchmark for vulnerability remediation, Argon ties for first place with a score of 68%, matching GPT-6 Astra, Grok 4.7, and Claude Opus 5.5. Google also reports that Argon outperforms its own Gemini 3.8 Flash Cyber model in vulnerability discovery, scoring 85.8% on Google's internal real-world vulnerability benchmark compared with 71.0% for Flash Cyber.
Early testing suggests practical impact. Wiz, a cloud security company, is already using Argon through its Scan for Good initiative, which protects critical public infrastructure. In an early demonstration, Argon uncovered a critical vulnerability that exposed sensitive personal information across healthcare software used by hospitals worldwide, a severe risk that previous frontier models had missed. Google says that for trusted defenders and its own internal teams, Argon is being released without cyber guardrails so they can use its full frontier-level cybersecurity capabilities.
Access is being managed through the Fairwind Program, which Google launched in September to give governments and vetted organizations early use of advanced cyber defense tools. Fairwind partners, which number more than 650 globally, must agree to strict operational standards: limiting access to internal cybersecurity, incident response, or penetration testing teams, and deploying protections like multi-factor authentication. Google conducts background checks on applicants and does not allow partners to share, redistribute, or sell access to the model.
How Google is already using Argon
Argon is not just a research product. Google says thousands of its engineers already use the model for daily work, from debugging to large-scale codebase migrations. The company cites several concrete examples that give a sense of what the model can do in practice. A team of Argon agents analyzed fleet-wide profiling telemetry to identify and apply memory optimizations across Google's data centers, freeing up over 300 TiB of memory once rolled out, with estimated total savings of 500 TiB to 1 PiB. Argon agents are also working on migrating C/C++ codebases to Rust across Google, scaling from tens of thousands of lines in core libraries like re2 and libgav1 up to more than 800,000 lines for the Fuchsia Zircon kernel.
One of the more striking internal results came from the video decoder library libgav1. Argon agents took an existing Rust port and replaced 32K lines of SIMD code by running many rounds of profile-guided experiments, studying the compiler's output, and writing safe Rust that the compiler would vectorize automatically. The result is a memory-safe video decoder that runs 2.7x faster than the Rust port, with identical video output, bringing it closer to the optimized C++ version. Google says a separate team used Argon to help quantum computing researchers optimize the spacetime resources of subroutines, beating a published baseline by 40% in a matter of minutes.
Pricing and rollout timeline
Google has announced introductory API pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% below the standard input rate. After the introductory period, pricing will rise to $4 per million input tokens and $20 per million output tokens, though Google has not specified when the increase takes effect. That eventual pricing is roughly comparable to Anthropic's Claude Opus 5.5 and roughly half of OpenAI's GPT-6 Astra, which costs $10 per million input tokens and $50 per million output tokens.
Availability will be phased. The Fairwind Program gets access first. Google says wider access will start with paid API customers and Google AI Ultra subscribers after an initial testing phase, but it has not given a specific date for general availability. The company says it is actively engaged in the U.S. government's voluntary process for pre-release model access while it iterates on guardrails and gathers feedback from early testers.
The competitive picture
The launch lands in a crowded and fast-moving field. OpenAI released GPT-6 Astra in early September with similar claims about coding, computer use, and cybersecurity, and Anthropic has been rolling out its Mythos-class models with strict access controls. All three labs are now following a similar playbook: release a frontier model, tout benchmark wins, and carefully gate the most dangerous cybersecurity capabilities behind trusted access programs. The differences are in emphasis and execution.
Google's decision to put cyber defense at the center of the launch is notable. It reflects both a genuine capability improvement and a strategic bet that defenders need frontier AI help sooner than general consumers. Analysts caution that benchmark scores should be treated as hints, not proof, and that enterprises should evaluate Argon on their own workloads rather than switching solely on vendor claims. The key metric, as one analyst put it, is not cost per token but cost per successful business outcome.
What to watch for
For developers and enterprises, the most important near-term question is how quickly Argon moves beyond the Fairwind Program. Google's stated roadmap is clear, but the timeline remains open. In the meantime, the model's 1 million token output limit and strong performance on long-horizon tasks suggest that the next phase of AI competition is less about raw intelligence and more about sustained, reliable execution over complex workflows. If Argon holds up to real-world use, Google may have reclaimed the frontier lead it appeared to lose over the summer. For now, it is a powerful tool that most of us can only read about.