Artificial Intelligence
Google announces Gemini 4 Argon rollout and pricing
Google said Gemini 4 Argon was rolling out first to trusted cyber defenders, with broader access later, while it set token pricing and promoted benchmark and internal performance claims.
- Google announced Gemini 4 Argon on 30 September 2026 and said it was rolling out first to trusted cyber defenders through its Fairwind Program.
- The company said it was using a phased release, including the U.S. government’s voluntary pre-release model access process, before broader access to developers, enterprises and consumers.
- Google said Argon would start at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off.
- It also said the model had a 1M-token output limit, up from 64K, and that it had shown benchmark gains in coding, legal, finance, long-video understanding and vulnerability remediation.
- Google said Wiz had already used Argon through Scan for Good and that an early demonstration uncovered a critical vulnerability in healthcare software used by hospitals worldwide.
Google said Gemini 4 Argon was its new frontier model and that it was starting to reach a small group of trusted cyber defenders through the Fairwind Program. It said wider availability would follow for developers, enterprises and consumers.
The company said it was taking a staged approach while it worked with the U.S. government’s voluntary pre-release model access process and collected feedback from early testers. It also said Argon would launch at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% less than standard input.
Google said Argon was already being used inside the company for coding, research, writing and other work. It claimed results including 77.9% on DeepSWE v1.1, a leading position on the Vals Index, and a 91.7% score on LVBench.
On cybersecurity, Google said the model could find, check and patch vulnerabilities, and that it would be released without cyber guardrails to trusted defenders and internal teams. It said Wiz had used the model through its Scan for Good initiative and that an early test had found a serious flaw in healthcare software used by hospitals worldwide.
Google said it was also strengthening safeguards before broader release, including work on misuse detection, prompt-injection defences, misalignment monitoring and hardened sandboxing. It said those measures had been tested by internal and external red teams using manual and automated attacks.
Named in this story
Companies
- announced Gemini 4 Argon and described its rollout and uses
- Wiz
- was already using the model for cybersecurity defence
- Zapier
- was named as the source of AutomationBench
Organisations
- Fairwind Program
- the programme through which the first rollout was made
Governments and agencies
- U.S. government
- was said to be involved in a voluntary pre-release model access process
Products and systems
- Gemini 4 Argon
- the new frontier model being launched
- Scan for Good
- Wiz’s initiative for finding and remediating exposures
- DeepSWE v1.1
- the benchmark on which Google said the model scored 77.9%
- Vals Index
- the benchmark Google said the model led on across economic work
- Vals Finance Agent v2
- the benchmark for multi-step financial research
- Harvey’s Legal Agent Benchmark
- the benchmark for legal research and drafting
- AutomationBench
- the benchmark for end-to-end execution across business functions
- LVBench
- the benchmark for long video understanding
- CWE-bench v1
- the benchmark for vulnerability remediation
- 3.8 Flash Cyber
- the earlier model Argon was compared against
- Gray Swan’s Indirect Prompt Injection (IPI) benchmark
- the benchmark used to measure prompt-injection robustness
- Frontier Safety Framework
- Google’s framework for handling misuse risks
- Google AI Ultra
- the subscription tier that would get access after launch
How the source tells it
The piece read as promotional launch copy, with repeated self-ranking and safety assurances around a new model release.
- vendor boosterism repeated superlatives and self-reported benchmark scores were used to present the model as leading
- reassurance the launch was wrapped in phased access, red-teaming and safeguard language