Google unveiled Gemini 4 Argon on Wednesday, introducing its new frontier model with benchmark results that place it ahead of rival AI systems on cybersecurity tests. The model scored 77.9% on DeepSWE v1.1 and posted a 0.7% attack success rate on Gray Swan's prompt injection benchmark, ahead of Claude Opus 5.5 and Claude Fable 5.1, which both scored 1.0%.

The launch arrives one week after the release of Claude Opus 5.5 and one day after GPT 6.1 Sol, the latest in a string of rapid-fire frontier model releases from leading American AI labs — a cadence that has made launch-day benchmark charts the industry's running scoreboard.

According to Google's official announcement,on is the company's most capable model to date, built for coding, office work and cyber defense.

Strong Software Engineering Scores, With Caveats

DeepSWE v1.1 measures whether an AI model can complete long, messy, real-world software engineering jobs, with results scored as a percentage. Argon hit 77.9%, ahead of Claude Opus 5.5 at 74.2%, GPT-6 Astra at 74.1% and Claude Fable 5.1 at 67.4%. For scale, Gemini 3.6 Flash managed 49% on the same test in July.

The model can also write up to 1 million tokens in a single reply, up from 64,000. A token is a chunk of text roughly three-quarters of a word, putting the new ceiling at about 750,000 words versus roughly 48,000 previously — headroom aimed at the long-form coding and office work the model is built for.

The numbers come with a caveat: Google computed its own DeepSWE score, while the figures for rival models came from a public leaderboard and company reports. Google's comparison table also concedes ground. Argon leads on 12 of 18 benchmarks, ties one and trails on five, a mix of coding, science and computer-control tests — a blend of self-run and third-party methodologies worth keeping in mind when scanning headline numbers.

Cybersecurity Takes Center Stage

The model's headline capability is cyber defense. In an indirect prompt injection attack, a secret instruction is hidden inside an email or other content; when an AI assistant reads it, the test is whether the model obeys the stranger instead of its user. Such attacks are a nightmare scenario for anyone handing an AI access to their inbox or shopping cart.

On Gray Swan's Indirect Prompt Injection benchmark, which hides malicious instructions in content that AI agents read and scores how often the attacks succeed within 15 tries, Argon landed at 0.7% — and lower is better. Claude Opus 5.5 and Claude Fable 5.1 both scored 1.0%. GPT-6 Astra came in at 8.5%. Grok 4.6 and Kimi K3 were tricked just over half the time, at 51.8% and 52.7%. The 52-point spread between the best and worst defenses on the list shows how uneven prompt-injection resistance remains across the industry.

Restricted Rollout Through the Fairwind Program

Argon goes first to vetted security teams through the Fairwind Program, Google's limited-access cyber defense initiative, which launched September 2 with more than 650 partners, including governments and critical infrastructure operators. The model ships "without cyber guardrails" — the built-in refusals that normally stop a model from helping with hacking.

The logic behind the move is that defenders need a model that can think like an attacker in order to patch holes before criminals find them. The catch is that the same skill cuts both ways, and Google says a phased rollout is the only safe path. The company is also taking part in the U.S. government's voluntary process for pre-release model access.

Google is not the first to place a cyber model behind a velvet rope. An early version of Anthropic's Claude Mythos helped find 271 vulnerabilities in Firefox — 271 security holes Mozilla then patched. OpenAI has taken a similar route with its Trusted Access for Cyber program, making restricted cyber access an emerging pattern among frontier labs.

Argon's cyber scores also jump over Gemini 3.8 Flash Cyber, the restricted model Google launched alongside Fairwind. On the Wiz Penetration Test Benchmark, an internal Google test that asks an AI to write working exploits against real web-application flaws without seeing the code and scores the share solved on the first try, Argon hit 70.9% against 58.2%. Google also says Argon helped security firm Wiz find a critical flaw in healthcare software used by hospitals worldwide — one that earlier frontier models had missed.

A Launch After a Rough Summer

The release follows a difficult stretch for Google. In July, the company shipped smaller Flash models but skipped the promised Gemini 3.5 Pro, and Alphabet shares fell about 4.4%. Argon also landed the same day President Trump unveiled a voluntary, penalty-free AI accord that Google's leadership signed.

Google says a wider release will come as soon as possible, starting with paid API customers and Google AI Ultra subscribers. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Google has not said when that period ends, only that standard rates are $4 and $20 — exactly double the introductory rates. The timing of the wider release and the length of the introductory window are the two dates to watch as Argon's rollout unfolds.