Google Launches Gemini 4 Argon, Claims Top Marks in Coding and Cybersecurity Benchmarks

Alphabet unveiled Gemini 4 Argon on Wednesday, describing it as the company's most advanced AI model to date. The release is limited at launch: Google is rolling out Argon first to select cybersecurity partners rather than through a general API or consumer release.

Google says the model sets a new record in real-world software engineering performance, ties for first place on a cybersecurity evaluation benchmark, and leads a separate benchmark measuring performance on professional tasks spanning finance, legal work, and other knowledge-intensive domains. The company did not disclose specific benchmark scores, context window size, or pricing in its announcement.

According to Google, Argon ties with OpenAI's GPT-6 Astra and xAI's Grok 4.7 on cybersecurity evaluation benchmarks, and outperforms both GPT-6 Astra and Anthropic's Fable 5.1 on the Vals Index, a separate professional-task benchmark. These are Google's claims; none of the competing labs have independently confirmed the comparisons.

The company says Argon is already deployed internally, where it has been used to optimize memory allocation across Google's data centers, freeing up hundreds of terabytes of memory without purchasing additional hardware, according to Google. Quantum computing researchers have also used the model, the company said, though no further detail was provided on those applications.

Phased rollout tied to safety review

Rather than a broad release, Google is starting with trusted cybersecurity partners while working with the U.S. government on pre-release safety evaluations. Tulsee Doshi, Google's Gemini model product lead, told CNBC the staged approach reflects the model's dual-use profile in cyber defense.

"Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible," Doshi said.

Google said Argon marks a significant capability jump from Gemini 3.8 Flash Cyber, a smaller cybersecurity-focused model the company released earlier this month, particularly in vulnerability discovery. Before a broader public launch, Google says it is scaling safeguards in four areas, including misuse prevention and resistance to prompt injection attacks.

Context: timing and industry backdrop

The launch comes one day after CEO Sundar Pichai signed a voluntary AI safety accord with President Donald Trump, following a White House meeting with other major tech executives addressing AI safety concerns.

Gemini 4 Argon arrives roughly a year after Gemini 3, which Google credited with restoring its competitive position in the frontier model race. The Argon release follows a period in which Google had emphasized faster, lower-cost Flash-tier models rather than new frontier releases.

What this means

Google's decision to gate Argon behind a cybersecurity-partner rollout rather than a standard API launch signals the company treats the model's dual-use capabilities — strong vulnerability discovery paired with frontier coding performance — as a genuine risk factor, not just a marketing framing. The comparisons to GPT-6 Astra, Grok 4.7, and Fable 5.1 are notable, but all figures come from Google's own testing; independent verification from AI21, Epoch AI, or similar third-party evaluators has not yet surfaced. Until Google publishes a model card with concrete benchmark numbers, context window specs, and pricing, Argon's actual standing versus rival frontier models remains a claim to track rather than a settled fact.