Gemini 4 Argon goes to cyber defenders first, without cyber guardrails, and with no public API date

Google announced Gemini 4 Argon on September 30, and almost nobody can call it yet. The announcement post, written by Koray Kavukcuoglu, says Argon "is rolling out to a set of trusted cyber defenders through our Fairwind Program," with developers, enterprises and consumers to follow "as soon as possible." It has published prices but no date for general access. As of October 6, Argon doesn't appear on Google's Gemini API models page, pricing page or changelog.

Defenders first, and without the cyber guardrails

The order of the rollout is deliberate. Google writes that "safely releasing frontier capabilities at this level requires a phased approach," and says it is "actively engaged in the U.S. government's voluntary process for pre-release model access." The early cohort also gets a different model configuration than everyone else eventually will:

For trusted defenders and our own internal teams at Google, we'll be releasing Argon without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities.

Google, Gemini 4 Argon announcement

The public version is meant to refuse. Google says the model "is designed to refuse harmful requests while preserving legitimate, dual-use scientific research," and that it is working on "techniques to monitor the model's internal activations to spot misuse." Two populations, two sets of behavior: the vetted few get the unrestricted model first.

On security evidence, Google cites CWE-bench v1, where Argon "ties for first place with a top score of 68%." The post doesn't say which model it ties with. It also cites internal comparisons against 3.8 Flash Cyber, the Fairwind-only variant released in September, on Google's internal vulnerability benchmark and on Wiz's black-box penetration testing benchmark, but it publishes no scores for either.

The 1M-token output limit

The largest specification change is output length. Google is "significantly expanding the model's output token limit to an industry-leading 1M tokens, up from the previous 64K tokens," to make room for trajectories of "hundreds of thousands of tokens."

Pricing makes that a real budget question. Argon launches at an introductory "$2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price." A footnote adds that after the introductory period, "the price of $4 per 1M input tokens and $20 per 1M output tokens will apply." The post doesn't say when the introductory period ends. At those rates, a single response that uses the full output limit would cost $10 now and $20 later, before any input charge.

The other numbers

Google reports a state-of-the-art 77.9% on DeepSWE v1.1, 91.7% on LVBench (long video), and 51.3% on Zapier's AutomationBench, where it says Argon ranks first. The internal examples are specific. Argon agents freed "over 300 TiB of memory once rolled out" across Google's data centers. They replaced 32K lines of SIMD code in a Rust port of the libgav1 video decoder, producing a version that "runs 2.7x faster than the Rust port." C/C++-to-Rust migrations are reaching "800K+ lines for the Fuchsia Zircon kernel," which Google says are "undergoing rigorous automated and manual auditing, emulation testing, and review before rolling out to production." None of these figures has been independently reproduced yet.

Monitoring the chain of thought

One safeguard section stands out. Google says it deploys "misalignment mitigations that monitor Argon's chain-of-thought and actions and stop execution when necessary." It used a similar monitor during training, "taking careful precautions against feeding the findings back into training so as to not risk shaping Argon's reasoning to evade our monitoring." The post then asks the rest of the industry "to preserve reasoning transparency." For developers, that means the shipped model's execution may be halted by a monitor reading its reasoning. The post doesn't say how such a stop will appear in the API.

For now there's no model ID to code against and no availability date. The paid API customers and Google AI Ultra subscribers named as the next in line will have to wait.

Primary sources: Google, "Gemini 4 Argon", Gemini API models, Gemini API pricing, Gemini API changelog, read 2026-10-06.