Archive · 1 Oct 2026
Google Launches Gemini 4 Argon for Cybersecurity and Coding
Google has made a new artificial intelligence model called Gemini 4 Argon.
It is meant to help with coding, research, writing, business tasks and understanding images and videos.
Google says Argon is especially good at finding and fixing weaknesses in computer software.
The model can also check whether a suspected weakness is real before suggesting a repair.
For now, only selected cybersecurity partners can use it through the Fairwind Program.
Google employees are using it for debugging, code migrations, quantum research and other large projects.
Argon can produce up to one million tokens, helping it work on very long tasks.
Google says it performed very well in tests, but more independent testing is needed to know how safe and dependable it is.
Google unveiled Gemini 4 Argon on September 30, 2026, as a frontier AI model for coding, research, enterprise work, writing and visual analysis.
Argon is initially available to trusted cybersecurity partners through Google’s Fairwind Program, with paid API users and Google AI Ultra subscribers expected to receive access later.
Google says Argon can discover, verify and repair software vulnerabilities with limited human intervention, including a flaw in healthcare software that other frontier models reportedly missed.
The model has a one-million-token output limit and is being used internally for debugging, code migration, quantum research and data-centre memory optimization.
Google reported leading benchmark results, but broader independent testing is needed to assess Argon’s reliability, safety and cybersecurity risks.
- Who
- Google introduced Gemini 4 Argon and is initially providing it to trusted cybersecurity partners.
- What
- A frontier AI model for long-running coding, research, enterprise, visual-analysis and cybersecurity tasks was launched.
- When
- September 30, 2026.
- Where
- Argon is being distributed through Google’s Fairwind Program and used internally by Google engineering and research teams.
- Why
- Google says Argon is intended to handle complex workflows and improve defensive cybersecurity, while the company competes with other AI developers.
Google’s Claims and Release Strategy
Safety and Reliability Cautions
Cybersecurity automation
Google’s Claims and Release Strategy
Google says Argon can autonomously find, verify and patch important software vulnerabilities, potentially accelerating defensive security work.
Safety and Reliability Cautions
The model’s cybersecurity claims come mainly from Google’s evaluations and early use, while broader testing must determine whether it reliably fixes real-world flaws without creating new risks.
Benchmark performance
Google’s Claims and Release Strategy
Google reports top or leading results across software engineering, business, legal, finance, video and vulnerability-remediation benchmarks, including a 77.9% DeepSWE score.
Safety and Reliability Cautions
Benchmark results can depend on test design, scoring methods and the versions of competing models used, so they do not by themselves establish overall superiority.
Access and safeguards
Google’s Claims and Release Strategy
Google is limiting initial access to trusted cyber defenders, monitoring model behavior and strengthening protections against misuse, prompt injection and harmful cyber or CBRN requests.
Safety and Reliability Cautions
Argon’s advanced capabilities and testing without usual cyber guardrails create risks that require careful assessment before wider availability; broader concerns about increasingly autonomous AI have also prompted calls for slower development.
- Model
- Gemini 4 Argon
- Initial access
- Trusted cyber defenders through Google’s Fairwind Program
- Planned wider access
- Paid API customers and Google AI Ultra subscribers are expected to receive access before a broader rollout
- Cybersecurity capability
- Discovering, validating and repairing software vulnerabilities with limited human intervention
- Output limit
- One million tokens, up from 64,000 tokens for the previous-generation Argon
- Reported benchmarks
- 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, 91.7% on LVBench and 68% on CWE-bench v1
- Launch pricing
- $2 per million input tokens and $10 per million output tokens; cached input tokens receive a 95% discount
