Gemini 4 Argon: Google's comeback model with cyber defense and 1M-token output

Google announced Gemini 4 Argon on 30 September 2026, and close to a year had passed since Gemini 3, the company's previous flagship generation. The pitch comes in three lanes: real-world software engineering at production scale, enterprise knowledge work such as legal and finance, and cyber defense, where Argon is supposed to find, validate and patch vulnerabilities on its own. The release itself is the most telling part. Instead of a public launch day, Argon goes out first to a small group of trusted cyber defenders through the Fairwind Program, while the model also runs through the voluntary US government process that lets agencies look at frontier models before the public does. Developers, enterprises and consumers wait, with no date attached. The pricing, however, is already set: two dollars per million input tokens and ten dollars per million output tokens during the introductory period, then four and twenty, with cached input tokens 95 percent off. The output limit rises from 64,000 to one million tokens, roughly fifteen times more headroom for a single reasoning trajectory.

The order of the rollout is a statement in itself. Cyber capability is double edged: the knowledge that closes a hole is the same knowledge that would open one. Google therefore removes its brakes only inside a shielded circle. Argon ships first without cyber guardrails to defenders in the Fairwind Program, an initiative that scans critical public infrastructure for free and accompanies the remediation of high-risk exposures, and to Google's own internal teams. Everyone else starts with a guarded version, and the circle widens to paid API customers and Google AI Ultra subscribers before developers, enterprises and end users follow. The official reasoning: releasing capabilities at this level safely requires a phased approach, built around feedback from early testers and guardrails tightened between stages. Read plainly, that also concedes the risks are not yet considered fully managed. A company in a hurry would simply have shipped.

On real engineering work, Argon reaches 77.9 percent on DeepSWE v1.1, a benchmark for long-horizon, real-world software tasks, which Google presents as a new state of the art. The Vals Index, which measures economic impact across finance, coding, legal and tax with each sector weighted by its contribution to US GDP, also goes to Argon, along with Vals Finance Agent v2 for multi-step financial research and Harvey's legal agent benchmark for research and drafting. On AutomationBench, Zapier's test of end-to-end execution of core business functions, Argon leads with 51.3 percent, a number worth reading as a ceiling: the best available model completes roughly half of those workflows in full. Long video understanding goes the same way, with 91.7 percent state of the art on LVBench. For the cyber lane, CWE-bench v1 tests vulnerability remediation, and Argon ties for first place at 68 percent on a leaderboard that also lists Grok 4.7, GPT-6 Astra and Claude Opus 5.5. The score builds on Gemini 3.8 Flash Cyber, which held frontier performance on the previous version of that benchmark.

The first field evidence comes from Wiz. The cloud security company runs Scan for Good, a program that protects critical public infrastructure for free by finding high-risk exposures and getting them fixed. In an early demonstration, Argon uncovered a critical vulnerability in healthcare software used by hospitals worldwide, exposing sensitive personal data, which previous frontier models had missed according to Google. On Google's internal vulnerability benchmark, Argon found exposures across codebases written in twenty programming languages. In Wiz's internal black-box testing, where a model sees only the externally visible attack surface of live systems and no source code, Argon reportedly beat its predecessor at mapping the attack surface, at identifying real vulnerabilities and at producing proof-of-concept evidence. The core claim of the lane is autonomy across the full loop: find, validate, patch. That claim is exactly why the release is phased, because the loop is tool equivalent to what attackers do.

Argon is already doing productive work at Google, with thousands of employees on specialized coding, deeper research and writing quality, per the announcement. Three cases show the shape. Quantum research: Argon optimizes subroutines for spacetime resources, qubits times gates, the resource pair that bottlenecks important applications, and in one case beat the published baseline by 40 percent in minutes rather than months. Memory: teams of Argon agents chewed through fleet-wide profiling telemetry from Google's data centers, autonomously identified and applied memory optimizations, and the rollout is expected to free over 300 TiB of memory, with total savings estimated between 500 TiB and 1 PiB. Migrations: Argon agents move C and C++ codebases to Rust, from tens of thousands of lines in core libraries such as re2 and libgav1 up to the Fuchsia Zircon kernel with more than 800,000 lines, each change passing automated and manual auditing, emulation testing and review before production. The libgav1 case is the sharpest detail: 32,000 lines of hand-tuned SIMD code were replaced with safe Rust written so the compiler adds vectorization automatically, and the result is a memory-safe video decoder running 2.7 times faster than the previous Rust port, with identical output.

The safety section of the announcement names four fronts. Misuse first: Argon is trained to refuse requests that would help with cyber attacks or with chemical, biological, radiological and nuclear applications, while legitimate dual-use research stays permitted under Google's frontier safety framework. Google also improves techniques that watch the model's internal activations and runs internal and external red teams with manual and automated attacks. Second, prompt injection, the hidden instructions inside web pages and documents that try to hijack the model: Google calls Argon its most resilient model yet against those attacks, cites automated red teaming and adversarial training, and leads Grey Swan's indirect prompt injection benchmark. Third, misalignment, the case of a model overshooting the user's intent: monitors watch the chain of thought and the actions and can stop execution, and a similar system already ran during training, alerting a dedicated incident response team. Findings from those monitors were deliberately not fed back into training, so the model cannot learn to shape its reasoning past them, and Google publicly urges the industry to preserve reasoning transparency. Fourth, environments: sandboxes are isolated and sealed before high-risk training or evaluation, with the practices to be shared with partners.

The delay is the context that gives this release its weight. In May, Sundar Pichai said the next major model would come in June, observers expected a Gemini 3.5 Pro, and neither arrived. By July, Axios reported on morale problems inside Google DeepMind, and NPR ran a longer piece asking how Google fell behind in the AI race. Meanwhile competitors pressed the top end: Anthropic shipped Fable in June, OpenAI launched GPT-6 Astra on 3 September and called it its best model yet. Bloomberg reported on release day, citing unnamed sources, that some Google employees found Argon's performance lacking in internal testing; Google rejected that characterization. Tulsee Doshi, head of Gemini products, told Axios that Argon is a well-rounded model with frontier capabilities across several domains and that Googlers had relied on it for weeks for their hardest coding and research problems. Koray Kavukcuoglu, SVP at Google DeepMind and chief AI architect, framed the model in the announcement as "our next era of frontier intelligence". Scale context: the Gemini app passed a billion monthly users in August, and ChatGPT reported a billion as well. If the benchmarks hold up, that would be a massive catch-up for a company many had written off.

Explicit limits: every benchmark figure above comes from Google's own reporting, independent verification does not exist yet while the market waits. The Bloomberg line rests on unnamed sources, and Google's denial is equally a claim, neither side publishes its evidence. Benchmarks say little about day-to-day feel; developer work on the released model will. The rollout plan can change before broad access, and the introductory prices are declared temporary. Status of this post: 2 October 2026.

Sources, all retrieved 02.10.2026:
- Google Blog: Gemini 4 Argon: our next era of frontier intelligence (30.09.2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- CNBC: Google rolls out Gemini 4 Argon, its most advanced AI model (30.09.2026): https://www.cnbc.com/2026/09/30/google-gemini-4-argon-ai.html
- Google DeepMind: Fairwind Program: https://deepmind.google/fairwind-program/
- Wiz: Scan for Good: https://www.wiz.io/scan-for-good
- Vals Index: https://www.vals.ai/benchmarks/vals_index
- CWE-bench v1: https://cwe-bench.com/
- YouTube (topic trigger): AI Revolution, Google Just Dropped Argon (02.10.2026): https://www.youtube.com/watch?v=b0Fg4riyZF0

About the author: I am Marcel, a graphic designer and the operator of the HUNTER cyberdeck, an agent setup on a used Google Pixel 6a. This post draws on Google's official announcement of 30 September 2026 and on coverage by CNBC, Axios, NPR and Bloomberg as listed above, status 02.10.2026. Last fact-checked on 02.10.2026.