The Gemini 4 era has begun.
Google on Wednesday announced a new AI model named Gemini 4 Argon — the first of the 4 series — which the company is calling its “next era of frontier intelligence.”
Google said the new model offers frontier performance in areas such as coding, cybersecurity defense, and enterprise knowledge work in domains like legal and finance.
Gemini 4 Argon is Google’s latest attempt to reassert itself in the race to build AI for high-stakes work, including security and coding. While Google has launched several updates to its Flash series of models, it has fallen behind the frontier to rival models from OpenAI and Anthropic in areas such as engineering and cybersecurity.
The new model could put Google back on the frontier. The company previously promised to release a new frontier model, Gemini 3.5 Pro, in June. The model, which was postponed several times internally due to poor performance, was ultimately not released.
To begin with, Google is making Argon available to a set of trusted partners in its Fairwind Program, an initiative for vetted governments and cyber authorities to proactively test new models and address cyber vulnerabilities.
Google said Argon ties with OpenAI’s GPT-6 Astra for the highest score on the CWE-bench, a test of how models can find and patch security vulnerabilities. Google also claims the model sets a new high score on a test of how models perform in real-world, long-horizon engineering tasks.
Google didn’t say when Gemini 4 Argon will be publicly available, but said it was “actively engaged” in the US government’s early-access framework for assessing cybersecurity and other risks of frontier models before launch.
Inside Google, employees are already using the new model for tasks such as debugging and large-scale codebase migrations, the company said on Wednesday.
It also said Argon has helped engineers free up more than 300 tebibytes of memory across its data centers without new hardware.
To reduce the risks of Argon acting in ways it shouldn’t — a behavior known as misalignment — Google said it implemented systems to track Argon’s chain-of-thought and stop it peforming when necessary.
Google said these safety measures also included precautions around giving the model feedback, to avoid teaching Argon to evade monitoring.
Have something to share? Contact this reporter via email at [email protected] or Signal at 628-228-1836. Use a personal email address and a non-work device; here’s our guide to sharing information securely.
