Google's Gemini 4 Argon opens up to cyber defenders first

Google announced Gemini 4 Argon on 30 September, and the first people to get their hands on it are not chatbot users but cyber defenders enrolled in the company's Fairwind Program. The model is rolling out to that trusted group first, while Google works through what it calls a phased release and gathers feedback from early testers.

A million tokens of headroom

The headline number is output length. Argon can generate up to 1 million tokens in a single trajectory, up from the 64,000-token ceiling Google shipped previously. The argument is that the headroom lets the model hold a long chain of reasoning together instead of losing the thread partway through a difficult problem.

Pricing is set at an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input rate.

What Google says it is already doing internally

Google says thousands of its own engineers and researchers are using Argon. In quantum computing, the company says the model beat a published baseline for optimising the spacetime resources (qubits multiplied by gates) of bottleneck subroutines by 40% in a matter of minutes. Agents built on Argon also analysed fleet-wide profiling telemetry and applied memory optimisations across Google's data centres, freeing more than 300 TiB once rolled out, with an estimated total saving of between 500 TiB and 1 PiB.

The same agents are working on C and C++ to Rust migrations. Google says the effort scales from tens of thousands of lines in core libraries such as re2 and libgav1 up to more than 800,000 lines for the Fuchsia Zircon kernel. For libgav1, its open source video decoder, Google says Argon took an existing Rust port and replaced 32,000 lines of SIMD code by running repeated profile-guided experiments and studying the compiler's output.

Benchmarks and the security pitch

Google reports 77.9% on DeepSWE v1.1, first place on Zapier's AutomationBench with 51.3%, 91.7% on LVBench for long-video understanding, and a tie for first on CWE-bench v1 with 68%. It also says Argon leads the Vals Index, which weights finance, coding, legal and tax work by each sector's contribution to US GDP.

For cyber defenders, Google is releasing Argon without cyber guardrails so they can use its full vulnerability-hunting capability. Wiz is already using the model through its Scan for Good initiative, and Google says Argon uncovered a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, a risk previous frontier models had missed.

Safeguards before a wider release

Google lists four areas it says it hardened before broadening access: refusing harmful requests while preserving legitimate dual-use research, resilience to indirect prompt injection, chain-of-thought monitoring to catch misaligned behaviour, and sealed sandbox environments for high-risk training and evaluations. Google says Argon will reach developers, enterprises and consumers as soon as possible, starting with paid API customers and Google AI Ultra subscribers.

Our opinion

The launch order is the story here. Handing a frontier model to cyber defenders before the public flips Google's usual consumer-first playbook, and the million-token output ceiling is a bet that long, autonomous work rather than chat is where model value now sits. The pricing sharpens that bet: $10 per million output tokens only looks cheap if Argon finishes the job in one pass, and Google's benchmark figures are still its own.