Gemini 4 Argon: Google Launches Its Frontier Model, First to Cyberdefenders

A launch in reverse

Google DeepMind presents Gemini 4 Argon as its new frontier model, built to sustain deep reasoning over long, complex tasks. Three domains are targeted: real-world software engineering, enterprise knowledge work (legal, finance), and cyberdefense.

The timeline is unusual. The model isn't going to developers first — it's going to a group of trusted cyberdefenders via the Fairwind program. Google says it is participating in the US government's voluntary pre-release model access process.

The general public, businesses, and developers will come next, "as soon as possible," starting with paying API customers and Google AI Ultra subscribers. No date is given. Introductory pricing, however, is announced: $2 per million input tokens, $10 for output, and 95% off cached input.

What Argon is already doing at Google

The post emphasizes internal use, with thousands of Googlers reportedly using it for specialized code, deep research, and writing. Three quantified examples are highlighted.

In quantum computing, Argon helps optimize the space-time resources (qubits × gates) of critical subroutines; in one case, it beat the published benchmark by 40% in a matter of minutes. On infrastructure, a team of Argon agents analyzed fleet-wide profiling telemetry to apply memory optimizations on its own: over 300 TiB freed, with 500 TiB to 1 PiB in expected savings.

The third project is migrating C/C++ code to Rust, from libraries like re2 and libgav1 to Fuchsia's Zircon kernel and its 800K+ lines. On libgav1, the agents replaced 32K lines of SIMD code with safe Rust that the compiler vectorizes on its own, producing a decoder 2.7x faster than the existing Rust port, with identical video output. Google notes these rewrites still go through automated and manual audits, emulation testing, and review before reaching production.

1 million output tokens

The most concrete technical change for users is the output limit, which jumps from 64K to 1M tokens. Google calls it the highest in the industry.

The idea: let the model produce hundreds of thousands of tokens in a single trajectory, solving a hard problem "in one go" rather than through repeated back-and-forth. It's a bet on long agentic tasks, and a potentially steep cost at $10 per million generated tokens.

Enterprise and multimodal: a flood of benchmarks

On code, Argon sets a new state of the art on DeepSWE v1.1 (77.9%), which measures long software engineering tasks. Outside of code, it tops the Vals Index, which weighs finance, code, law, and tax according to their share of US GDP, and claims strong results on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark.

On AutomationBench, Zapier's benchmark for end-to-end execution of business functions, Argon ranks first with 51.3%. On vision, it reaches 91.7% on LVBench, focused on long video understanding, and Google highlights chart analysis and taking action from a series of documents.

Cyberdefense, the core of the announcement

Argon was specifically trained for defense: finding, validating, and fixing critical vulnerabilities autonomously. For trusted defenders and internal teams, it will ship without cyber guardrails, to unlock its full capability.

Wiz is already using it in its Scan for Good initiative, which protects critical public infrastructure for free. The model spotted a critical vulnerability exposing sensitive personal data in healthcare software used by hospitals worldwide — a flaw that previous models had missed.

On CWE-bench v1, which evaluates vulnerability remediation, Argon ties for first place with 68%. Google also cites progress on its internal benchmark (vulnerabilities found across codebases spanning 20 languages) and on Wiz's black-box pentesting benchmark, compared to its predecessor 3.8 Flash Cyber.

Guardrails before the rollout

Google details four workstreams. Against misuse (cyber and CBRN), the model refuses dangerous requests while preserving legitimate dual-use research, and Google monitors the model's internal activations to catch misuse, with testing by internal and external red teams.

Against indirect prompt injection, Argon is said to be the most robust model on the market on Gray Swan's IPI benchmark, thanks to automated red teaming and adversarial training. For misalignment, Google monitors the model's chain of thought and actions and can halt execution.

Notably: alerts from training-time monitoring are not fed back into training, to avoid teaching the model to evade oversight. Google calls on the rest of the industry to preserve reasoning transparency. Finally, sandbox environments are isolated and sealed off before high-risk training and evaluations.