Key Highlights
- Alphabet unveiled Gemini 4 Argon on September 30, calling it its most advanced artificial intelligence model yet.
- Argon scored 77.9% on DeepSWE v1.1 and 68.9% on the Vals Index, leading both, according to Google.
- Access begins with trusted cybersecurity partners, with broader release phased in after safeguards are scaled in four areas.
- Argon agents freed more than 300 TiB of data center memory at Google, the company said.
A Frontier Model Aimed at Coding, Security and Professional Work
Alphabet (NASDAQ:GOOGL), the parent company of Google, unveiled Gemini 4 Argon on September 30, calling it its most advanced artificial intelligence model yet. The company said the model delivers major improvements in coding, cybersecurity and complex professional work, and that it handles long, multi-step tasks well.
Argon follows Gemini 3, which arrived nearly a year ago and put Alphabet back at the front of the AI model race. It also builds on Gemini 3.8 Flash Cyber, a security-focused model Google released in September. Google said Argon is a meaningful step up from that model, outperforming it in vulnerability discovery.
Google's product team described the new model as well-rounded and said a system at this level of frontier performance carries real importance for defenders working to protect software. The model also raises the output limit to 1 million tokens, up from 64,000 in earlier Gemini models.
Benchmark Results: Leads, Ties and Gaps
Google said Argon sets a new record in real-world software engineering. Its published comparison shows the model at 77.9% on DeepSWE v1.1, ahead of 74.2% for Anthropic's Claude Opus 5.5 and 74.1% for OpenAI's GPT-6 Astra.
In cybersecurity, the company said Argon ties for first. It scored 68% on CWE-bench v1, a vulnerability repair benchmark, level with GPT-6 Astra and Grok 4.7, while Claude Opus 5.5 posted 67%.
Argon also leads the Vals Index, which measures performance across finance, legal and other professional tasks. It scored 68.9%, compared with 67.0% for Claude Opus 5.5, 65.8% for Claude Fable 5.1 and 63.1% for GPT-6 Astra.
The results are not a clean sweep. On FrontierSWE v2, Argon scored 55.0%, behind 65.5% for GPT-6 Astra and 62.3% for Claude Opus 5.5. On Terminal-Bench 4.0, it scored 57.4%, behind 66.4% for Claude Opus 5.5. Most of these figures come from Google's own published comparison.
Early Internal Use at Google
Google said Argon is already in use inside the company. A team of Argon agents analyzed fleet-wide profiling data from its data centers and applied memory optimizations, freeing more than 300 TiB of memory without the purchase of additional hardware. Google projects total savings of between 500 TiB and 1 PiB.
Quantum computing researchers have also used the model. Google said it improved a published baseline for quantum subroutine resources by 40% in a matter of minutes.
Phased Rollout Starts With Cyber Defenders
Google plans to launch Argon in phases, starting with trusted cybersecurity partners through its Fairwind Program, while working with the US government on pre-release safety evaluations. The company said beginning this way gives it more confidence and puts a model trained for cyber defense in the hands of defenders as soon as possible.
Before a public launch, Google said it is looking to scale safeguards in four key cybersecurity areas, including misuse and prompt injection. On Gray Swan's indirect prompt-injection benchmark, Argon recorded a 0.7% attack success rate, compared with 1.0% for Claude Opus 5.5 and 8.5% for GPT-6 Astra.
Paid API customers and Google AI Ultra subscribers are next in line for access, and no public release date has been announced. Google has set introductory pricing at USD 2 per million input tokens and USD 10 per million output tokens.
The release came a day after Alphabet's chief executive signed a voluntary agreement with President Donald Trump and major tech executives at the White House. The signing followed a luncheon convened to address rising AI safety concerns.
What to Watch as Access Widens
Google has not announced a public release date for Argon, or release dates for other Gemini 4 models. Items to watch include progress on safeguards across the four cybersecurity areas, the outcome of the pre-release safety evaluations with the US government, whether independent testing reproduces Google's reported benchmark results, and how Argon's scores compare on the benchmarks where it currently trails.






Please wait processing your request...