Google Unveils Gemini 4 Argon With a Claimed Benchmark Lead, Cyber Defenders First

Google unveiled Gemini 4 Argon, the first model in its Gemini 4 family, on September 30, rolling it out first to trusted cyber defenders through its Fairwind Program, with broader access to follow for paid API customers and Google AI Ultra subscribers.

<blockquote><strong>Update, October 25:</strong> Argon began rolling out on October 13 to Google AI Pro and Ultra subscribers in the Gemini app and to the paid Gemini API, under the model ID gemini-4-argon, and has since reached AI Mode in Search for Pro and Ultra users (<a href="https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/">Google</a>, <a href="https://www.searchenginejournal.com/google-gemini-4-argon-ai-mode/592806/">Search Engine Journal</a>). Free users and AI Plus subscribers do not have it, and the introductory API price runs through November 30 (<a href="https://www.digitalapplied.com/blog/who-can-use-gemini-4-argon-every-google-plan">Digital Applied</a>).</blockquote>

What Google is claiming

Google describes Argon as a frontier model for real-world software engineering, enterprise knowledge work such as legal and finance tasks, and cybersecurity defence. Output length rises to 1 million tokens, up from 64,000 on the previous generation, aimed at long-running coding and analysis jobs.

Benchmark	Gemini 4 Argon (Google's figure)	Note
DeepSWE v1.1	77.9%	Software engineering
CWE-bench v1	68%	Vulnerability fixing; tied for first
AutomationBench	51.3%	Top spot, per Google
LVBench	91.7%	Long video understanding
Harvey Legal Agent Benchmark	19.6%	Rivals 3.8% to 5.4% (VentureBeat)
FrontierSWE v2	55.0%	GPT-6 Astra leads at 65.5%

All of these are Google's figures and have not been independently reproduced. VentureBeat reports that Argon leads or ties on 13 of the 18 benchmarks Google disclosed against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, with the widest gap on enterprise work, while Claude Opus 5.5 leads on terminal agent tasks. Its conclusion is that the right model still depends on the workload.

Axios notes that the launch came nearly a year after Gemini 3 and is widely read as Google's answer to OpenAI and Anthropic, and that Bloomberg found some Google employees disappointed with internal testing, a claim Google disputes. Tulsee Doshi, head of Gemini products, said: "Argon is a well-rounded model that has frontier capabilities across several domains."

Pricing and who can use it

Google charges an introductory $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached input, rising to $4 and $20 after November 30. VentureBeat says the standard rate matches Claude Opus 5.5 and sits well below GPT-6 Astra's $10 and $50. There is no free API tier.

Safeguards and the government process

Google says Argon was trained for cybersecurity defence and ships with safeguards against cyber misuse and against chemical, biological, radiological and nuclear misuse. It also describes monitoring of the model's reasoning and actions for misalignment, prompt injection defences tested on the Gray Swan benchmark, and secured environments for its highest-risk evaluations. Google is also taking part in the US government's voluntary process for pre-release model access, which helps explain why the launch started with such a narrow audience, in the same week the FTC opened an inquiry into AI agent risks.

What to watch next

Now that paid API access is open, watch for independent evaluations of whether the legal and finance gains hold up outside Google's chosen benchmarks, and how usage shifts when the introductory pricing ends on November 30.

Related on Gadgets365

• Google restricts free Gemini users to Flash-Lite
• Google's new Gemini agent gets its own work email address
• Mistral opens Large 4 preview