Gemini 4 Argon Released: One Million Token Output and a New Era of Frontier Intelligence

Google DeepMind has officially launched Gemini 4 Argon, its next-generation frontier model designed to sustain deep, continuous reasoning across complex, long-horizon professional workflows. Unlike previous releases that emphasized general-purpose capability, Argon is focused on three high-value use cases: real-world software engineering, enterprise knowledge work in legal and finance, and defensive cybersecurity. Initial access is being rolled out through the Fairwind program to trusted cyber defenders, with broader availability for developers, enterprises, and consumers to follow.

One Million Token Output: Turning "Think Deeper" Into a Product Capability

The most headline-grabbing specification in Argon is its output token limit, raised from the previous 64K to an industry-leading 1 million tokens. This isn't just a number game — when a model has enough "thinking room" to generate hundreds of thousands of tokens within a single trajectory, it can complete complex reasoning chains in a single interaction that previously required repeated splitting and stitching.

For tasks requiring long-horizon planning — such as large-scale code migrations, multi-step financial research, or end-to-end legal drafting — this ability to "think it all the way through in one go" means fewer context breaks and less state loss. Google describes this as "a new level of reasoning depth," essentially trading a longer output space for more complete problem-solving capability.

On pricing, Argon starts at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. For a flagship model with million-scale output capacity, this is a relatively restrained pricing strategy.

Tokens are the basic units a large language model uses to process text — roughly half a word in English or one character in Chinese. The output token limit determines how much content a model can "write" in a single inference pass. Previously, mainstream flagship models generally capped output between 4K and 64K tokens. For complex tasks requiring many intermediate reasoning steps, models often had to truncate their thinking before reaching the limit, with external systems stitching together results from multiple calls — a process prone to context loss and logical breaks. Raising the limit to 1 million tokens means the model can maintain a complete reasoning chain of hundreds of thousands of steps within a single forward pass, without external state management, theoretically reducing the information loss caused by "segmented thinking." The tradeoff is significantly higher compute cost, which Argon's $10/million output token pricing (roughly 5× the input price) partly reflects.

Google Is Already "Eating Its Own Dog Food"

Argon is not a paper model. Google reports that thousands of employees are already using it in daily workflows, spanning professional coding, deep research, and writing. Several internal case studies are particularly compelling:

Quantum Algorithm and Memory Optimization

In quantum computing research, Argon helped researchers optimize the space-time resources (qubits × gates) of subroutines that bottleneck key applications, improving on published baselines by 40% in a matter of minutes in one case. At the infrastructure level, a group of Argon agents analyzed fleet-wide performance telemetry, autonomously identified and applied memory optimizations, and are expected to free up 300 TiB to 1 PiB of memory across Google's data centers.

Large-Scale Migration from C/C++ to Rust

Even more technically significant is the code migration work. Argon agents are migrating Google's internal C/C++ codebases to Rust, ranging from tens of thousands of lines in core libraries like re2 and libgav1 up to over 800,000 lines in the Fuchsia OS Zircon kernel. Taking the open-source video decoding library libgav1 as an example: Argon agents conducted multiple performance-guided experiments, studied compiler output, and replaced 32K lines of SIMD code with safe Rust, ultimately producing a decoder that is memory-safe and 2.7× faster than the existing Rust port, with output fully matching the original.

Google emphasizes that large-scale rewrites of critical systems like these undergo rigorous automated and human audits, as well as simulation testing, before being pushed to production.

Rust is a systems programming language focused on memory safety. Its core "ownership system" statically eliminates at compile time an entire class of memory safety vulnerabilities — dangling pointers, buffer overflows, data races — that have historically accounted for a large share of high-severity CVEs in C/C++ codebases. Google, Microsoft, the NSA, and others have publicly recommended migrating critical infrastructure to memory-safe languages, but manual C/C++ to Rust migration is extremely time-consuming, especially for low-level performance code involving SIMD (Single Instruction Multiple Data) intrinsics, which requires deep understanding of the original algorithmic intent to translate equivalently. What makes Argon's libgav1 migration noteworthy is not just the code volume (32K lines of SIMD), but that the result simultaneously achieved both "memory safety" and "2.7× faster than the existing Rust port" — two goals that typically involve tradeoffs — suggesting the model is capable of actively conducting performance-guided optimization after deeply analyzing compiler output.

Coding, Finance, Law: Cross-Domain Enterprise Capability

Argon has set new records across multiple enterprise-grade benchmarks. It achieved a new high of 77.9% on DeepSWE v1.1, which measures real-world long-horizon software engineering tasks.

Beyond coding, Argon leads on the Vals Index — a benchmark weighted by contribution to U.S. GDP that measures the economic impact of financial, coding, legal, and tax work. It also leads on domain-specific evaluations including Vals Finance Agent v2 (multi-step financial research) and the Harvey legal agent benchmark (legal research and drafting). On Zapier's AutomationBench (end-to-end business task execution), Argon ranks first at 51.3%.

Also worth noting is its visual understanding advantage: on LVBench, which measures long-video comprehension, Argon sets a new best score of 91.7%, enabling professional chart analysis, detail extraction from long videos, and action-taking based on document series.

Cybersecurity: From Passive Defense to Autonomous Vulnerability Patching

One of Argon's key differentiators is defensive cybersecurity. Google specifically trained it with the ability to autonomously discover, verify, and patch critical software vulnerabilities. For trusted defenders and internal teams, Google will release a version with "no cybersecurity guardrails," enabling access to its full frontier-level security capabilities.

Security firm Wiz has already been using Argon through its Scan for Good program, which freely finds and fixes high-risk vulnerabilities in critical public infrastructure. In an early demonstration, the model identified a critical vulnerability that previous frontier models had failed to catch — a piece of medical software used by hospitals globally was leaking sensitive personal information.

On benchmarks, Argon ties for first place on CWE-bench v1 with a top score of 68% for vulnerability remediation capability, and outperforms the previous-generation 3.8 Flash Cyber across Google's internal comprehensive vulnerability benchmark covering 20 programming languages and on Wiz's black-box penetration testing benchmark.

CWE (Common Weakness Enumeration) is a software security defect classification standard maintained by MITRE, covering hundreds of common vulnerability patterns such as injection, out-of-bounds access, and race conditions. CWE-bench is an automated vulnerability remediation evaluation suite built on this taxonomy, requiring the model to autonomously locate the root cause of known defects in a given codebase and generate patches that pass a test suite. A score of 68% means the model can independently complete the full "discover → understand → fix" pipeline in more than two-thirds of cases, rather than merely hinting to humans that "something might be wrong here." Prompt injection is an attack vector unique to agent scenarios: malicious content is embedded in external data processed by the model (such as web pages, documents, or tool return values) in an attempt to hijack the model's subsequent behavior. For agents autonomously executing long-horizon tasks, resisting indirect prompt injection is a foundational security capability for maintaining trustworthy action.

Staged Rollout: Safety Guardrails First

Given Argon's powerful capabilities, Google has opted for a cautious, staged release path and has explicitly participated in the U.S. government's voluntary pre-release model access process. Before launch, Google reinforced frontier safety guardrails across four dimensions:

• Abuse prevention: Refuses requests related to cyberattacks and CBRN (chemical, biological, radiological, nuclear) harms, while preserving legitimate dual-use research, and identifies misuse by monitoring model internal activations.
• Prompt injection resistance: Argon is claimed to be the most resistant model to indirect prompt injection attacks currently available, leading on Gray Swan's IPI benchmark.
• Misalignment monitoring: Deploys misalignment mitigation mechanisms that monitor Argon's chain of thought and actions, with the ability to abort execution when necessary — and deliberately avoids feeding monitoring findings back into training to prevent the model from learning to evade detection.
• System hardening: Isolates and seals sandbox environments before high-risk training or evaluation.

Google has also called on the industry to maintain reasoning transparency at critical moments of capability leaps, keeping the model's "thought process" available for identifying and diagnosing misalignment issues.

Conclusion

Gemini 4 Argon sends a clear signal: competition among frontier models is shifting from "general-purpose chat capability" to "can the model independently complete tasks within real, long-horizon, high-value professional workflows." With million-token output, leading cross-domain enterprise benchmarks, and a dedicated investment in defensive cybersecurity, Argon looks more like a "heavy-duty productivity partner" for engineers, finance and legal professionals, and security teams. The staged, safety-first release cadence also reflects Google's effort to balance capability against risk. The model will be made available first to paid API customers and Google AI Ultra subscribers.