Google announces Gemini 4 Argon, starting with vetted cyber defenders

Google announced Gemini 4 Argon on September 30, 2026, in a post signed by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. Argon is Google's new frontier model, and for now it goes only to a set of vetted cyber defenders through the Fairwind Program. Developers, enterprises and consumers get access later, on a date Google has not given.

What is new in Gemini 4 Argon

The clearest technical change is output length. According to the Gemini 4 Argon announcement (opens in a new tab), the model's output token limit is now 1M tokens, up from a previous 64K tokens. The post does not say which earlier model the 64K figure belongs to. It also gives no input context window.

Google's argument for the higher ceiling is that the model can reason through hundreds of thousands of tokens in a single run and finish hard problems in one pass.

The benchmark results Google reports are its own figures:

• DeepSWE v1.1 (long-horizon software engineering): 77.9%, which Google calls a new state of the art.
• AutomationBench, Zapier's benchmark for end-to-end business tasks: 51.3%, ranked first.
• LVBench (long video understanding): 91.7%.
• CWE-bench v1 (fixing security vulnerabilities): 68%, tied for first place.
• Vals Index: Google says Argon leads it but prints no score.

On vulnerability discovery, Google compares Argon with its earlier 3.8 Flash Cyber model on an internal Google benchmark and on Wiz's internal black-box penetration testing benchmark. The post describes Argon as ahead on both without publishing numbers.

What Google says Argon did internally

The post lists several internal results. For quantum computing research, Argon beat a published baseline for subroutine resource use by 40% in one example. A group of Argon agents found memory optimizations across Google's data centers that freed over 300 TiB of memory, with an estimated 500 TiB to 1 PiB in total savings.

Google also has Argon agents porting C and C++ code to Rust, from libraries such as re2 and libgav1 up to the 800K+ line Fuchsia Zircon kernel. These rewrites are still going through automated and manual audits before production. In libgav1, a video decoder, the agents replaced 32K lines of SIMD code in an existing Rust port. Google says the result runs 2.7x faster than that port, with identical video output.

Price and who can use Gemini 4 Argon

Google set an introductory API price of $2 per million input tokens and $10 per million output tokens. Cached input tokens cost 95% less than the input price. A footnote says that once the introductory period ends, the price becomes $4 per 1M input tokens and $20 per 1M output tokens. The length of the introductory period is not stated.

Today, access runs through the Fairwind Program. Google DeepMind's Fairwind page says it prioritizes governments and national cyber authorities, critical infrastructure operators and core technology platforms. Applicants go through background checks. Approved partners may give Argon only to internal security, incident response or penetration testing teams. They must use phishing-resistant MFA and cannot resell access. Partners can also run Argon inside CodeMender, Google's code security agent. Google says trusted defenders and its internal teams get Argon without cyber guardrails.

The next group in line is paid API customers and Google AI Ultra subscribers. As of October 5, 2026, Argon does not appear on the Gemini API models page (opens in a new tab) or the Gemini API pricing page (opens in a new tab). No model ID has been published.

Who Gemini 4 Argon matters for right now

Security teams at governments, infrastructure operators and large platforms can apply through the Fairwind interest form. Everyone else has nothing to call yet. Teams that expect to compare Argon with their current model can prepare an evaluation set now, for example by testing prompts and outputs with promptfoo, so the comparison is ready when the API opens.

Budgets need care. The post-introductory rates are double the introductory ones on both input and output.

Safety work Google describes for Argon

Google lists four areas it is working on before wide release. For misuse, the model is meant to refuse cyber and CBRN attack requests, and Google monitors internal activations to spot abuse, with internal and external red teams testing those safeguards.

For prompt injection, Google says Argon leads Gray Swan's Indirect Prompt Injection benchmark, without giving a score. A misalignment monitor reads Argon's chain of thought and actions and can stop execution. Sandboxes used for high-risk training and evaluation are being isolated and sealed.

Google also says it is taking part in the U.S. government's voluntary process for pre-release model access.

Limits and open questions about Gemini 4 Argon

• No model card, API model ID, rate limits, input context size or modality list is in the announcement.
• Every benchmark figure comes from Google. Several claims (Vals Index, Vals Finance Agent v2, Harvey's Legal Agent Benchmark, Gray Swan) have no published score in the post.
• The release date for API customers and Ultra subscribers is not given, and neither are regions.
• How long the introductory price lasts is unknown.
• The page's AI-generated summary calls 1M tokens a "token limit", while the post body says it is the output limit. The body is the more specific source.