Google has launched its Gemini 4 Argon AI model, enhancing capabilities for programming and cybersecurity with a focus on complex task execution.
On September 30, Google revealed its latest AI model, Gemini 4 Argon, designed to tackle programming, corporate tasks, and cybersecurity challenges. This flagship model emphasizes the execution of complex, multi-step tasks.
Currently, Gemini 4 Argon is being tested by select cybersecurity experts through the Fairwind program, alongside participating in a voluntary pre-screening initiative with the U.S. government for advanced models. Following this testing phase, the model will be available first to paying API users and subscribers of Google AI Ultra, though no public release date has been announced.
Enhanced Token Output
The model's token output capability has been significantly increased from 64,000 to 1 million tokens, allowing for better handling of intricate tasks and the generation of extensive outputs continuously. In the DeepSWE v1.1 benchmark, which evaluates engagement with real software projects, Argon achieved a score of 77.9%.
Performance varies by task, with Argon scoring 55% in the FrontierSWE v2 benchmark, compared to 65.5% for GPT-6 Astra and 62.3% for Claude Opus 5.5. In corporate task assessments, it scored 65.4% in Vals Finance Agent v2 and 51.3% in Zapier's AutomationBench, excelling with a score of 91.7% in the LVBench test for lengthy video comprehension.
The API will initially be priced at $2 per 1 million input tokens and $10 per 1 million output tokens, with significantly lower costs for caching input at $0.10 per million tokens. After the introductory period, prices are expected to rise to $4 and $20, though the specific timing for this increase remains unclear.
Internal Applications at Google
Gemini 4 Argon is already in use by thousands of Google employees, including teams optimizing data center operations, which have freed up over 300 TiB of RAM. The company anticipates potential savings of 500 TiB to 1 PiB from wider deployment.
Additionally, Argon is aiding in the transition of code from C/C++ to Rust, handling projects ranging from extensive libraries to the Zircon kernel of the Fuchsia OS, which contains over 800,000 lines of code. In one instance, Argon improved the performance of a quantum algorithm optimization by 40% within minutes.
Cybersecurity Focus
Argon has been specifically trained to autonomously detect, verify, and fix software vulnerabilities. In the CWE-bench v1 benchmark, it achieved a pass@1 score of 68%, comparable to GPT-6 Astra and Grok 4.7. The company Wiz is currently testing Argon in its Scan for Good program, where it identified a critical vulnerability in hospital software that could have compromised sensitive data, a flaw overlooked by other models.
Before a broader rollout, Google intends to enhance safeguards against misuse and potential attacks by incorporating hidden instructions. A monitoring system has also been established to oversee Argon's operations and halt any tasks that exceed user intentions. In September, Google also launched Gemini 3.8 Flash for advanced programming tasks and a specialized version, Gemini 3.8 Flash Cyber, focused on vulnerability assessment.
During cybersecurity tests, Gemini models were reported to have accessed the systems of three real companies instead of simulated environments, with Google confirming that the model ceased operations autonomously after detecting an error, ensuring no harm was caused.
