Google launched the Gemini 4 Argon artificial intelligence model on Wednesday (30) in Mountain View, California, opening the Gemini 4 family with a focus on deep reasoning, software engineering, and defensive cybersecurity. The system expands the output ceiling to 1 million tokens per response, up from the previous limit of 64,000 tokens found in earlier releases.

The model is rolling out initially to authorized cybersecurity defenders participating in the Fairwind Program and internal engineering teams at the company. According to Koray Kavukcuoglu, senior vice president at Google DeepMind and chief AI architect at Google, safely deploying capabilities at this scale “requires a phased approach” before reaching external enterprise customers, developers, and subscribers.

Gemini 4 Argon launch details and pricing tiers

The company introduced the system with an initial access structure for developers using its programming interfaces and enterprise tools. Commercial availability will begin on API channels before expanding to subscribers of the Google AI Ultra plan.

• Product name: Gemini 4 Argon
• Developer: Google DeepMind
• Launch date: September 30, 2026
• Initial input token price: US$ 2.00 per 1 million tokens
• Initial output token price: US$ 10.00 per 1 million tokens
• Cached input discount: 95% off the standard input rate
• Standard post-introductory rates: US$ 4.00 per million input tokens and US$ 20.00 per million output tokens
• Maximum output capacity: 1,000,000 tokens in a single generation

Benchmark results and performance in coding evaluations

In industry evaluations, Gemini 4 Argon recorded a score of 77.9% on the DeepSWE v1.1 benchmark, which tests long-horizon software engineering capabilities. In cybersecurity patching, the system scored 68% on CWE-bench v1, tying for the top position in remediating reported software flaws.

The model also achieved 51.3% on Zapier’s AutomationBench, securing first place in end-to-end multi-step workflow execution across core administrative operations. On multimodal tasks requiring long video comprehension, Argon registered 91.7% on the LVBench test, analyzing visual information alongside document streams. In economic evaluations, the model ranked first on the Vals Index, which measures performance across finance, legal, tax, and coding tasks weighted by sector contributions to United States gross domestic product.

External evaluations present a more nuanced picture across specialized developer environments. Independent evaluations, including metrics from Artificial Analysis and benchmarks such as Terminal-Bench 4.0 and FrontierSWE v2, show that competing architectures such as OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 maintain advantages over Argon in selected agentic workflows.

Autonomous code rewrites and infrastructure savings inside Google

Google engineers have integrated Argon into internal infrastructure workflows to optimize compute performance and rewrite legacy codebases. Agents built on the architecture migrated codebases written in C and C++ to memory-safe Rust, ranging from libraries such as re2 and libgav1 to critical components in the Fuchsia Zircon kernel encompassing more than 800,000 lines of code.

On libgav1, the open-source video decoder developed by Google, Argon agents replaced 32,000 lines of SIMD code through automated profiling and compiler feedback, producing Rust code that compiles into a memory-safe binary that runs 2.7 times faster than earlier ports. In data center operations, autonomous agents analyzed profiling telemetry across server fleets, applying memory optimizations that freed over 300 TiB of RAM upon deployment, with company projections estimating long-term savings between 500 TiB and 1 PiB.

In quantum computing research, the model assisted in optimizing spacetime resource constraints measured by qubits multiplied by gates for performance-critical subroutines. In internal trials, Argon beat published baseline calculations by 40% within minutes.

Cyber defense testing through the Fairwind Program

Google designed the model to address security operations by enabling autonomous discovery, validation, and remediation of system vulnerabilities. To facilitate defensive research, Google confirmed plans to supply an unaligned version without cyber guardrails to trusted participants in the Fairwind Program and internal security teams, granting operators the latitude to execute live penetration tests and vulnerability validation.

Check out: Google Gemini breaches networks of three firms in safety test

The Fairwind Program, established on September 3, 2026, includes more than 650 organizations worldwide, spanning government bodies and critical public infrastructure operators. Cybersecurity firm Wiz has integrated Argon into its Scan for Good initiative to scan public infrastructure for security exposures. During preliminary deployments, the model detected a critical vulnerability in healthcare management software that exposed sensitive personal data across hospitals worldwide, uncovering an issue missed by prior models.

Kavukcuoglu explained that Gemini 4 Argon “delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.”

Frontier safeguards and pending rollout details

The release follows internal restructuring in Google’s artificial intelligence timeline. The company had internally halted the rollout of Gemini 3.5 Pro in the first half of 2026 due to performance considerations, subsequently deploying Gemini 3.8 Flash Cyber in early September 2026 before advancing the Argon release under market competition from OpenAI and Anthropic.

Follow: Gemini penetrates 3 corporate networks during Google testing

To reduce operational hazards, Google incorporated monitoring mechanisms targeting indirect prompt injection attacks and misalignment. “We are deploying misalignment mitigations that monitor Argon’s chain-of-thought and actions and stop execution when necessary,” the company said in an official statement detailing its Frontier Safety Framework.

Google has not yet announced the closing date for the introductory API pricing of US$ 2.00 per million input tokens and US$ 10.00 per million output tokens, nor has it published the exact release dates for general access across third-party developer platforms and individual consumers.