On 30 September, Google unveiled its Gemini 4 Argon artificial intelligence model. A small group of vetted cybersecurity experts were granted early access via the Fairwind programme, the company said.
Briefly about the main points
- Google unveiled the Gemini 4 Argon on 30 September.
- Access was initially granted to trusted cyber security specialists.
- Next, they plan to make access available to paying API clients and Google AI Ultra subscribers.
- Bloomberg reported internal doubts regarding the coding, whilst Google denied this.
Who will be granted access, and when?
Google plans to gradually expand access to Gemini 4 Argon for developers, businesses and consumers following further testing. The company named paid API clients and Google AI Ultra subscribers as the first in the following categories.
Google has not announced a specific date for the wider launch. At this stage, the announcement does not mean that the model is fully available to all users.
What Google says about the model’s capabilities
The company is positioning Gemini 4 Argon for long-running and complex tasks in software engineering, finance, law and cybersecurity. Google has also reported strong results for the model in standardised tests, including 77.9% on DeepSWE v1.1, 51.3% on AutomationBench and 68% on CWE-bench v1.
Benchmarks are standardised tests used to compare models. High scores in these tests do not prove equivalent performance in every real-world scenario. Google also stated that Argon outperformed OpenAI Astra in a single cybersecurity skills test, but there is no independent verification of this particular comparison.
According to Google, the model can generate up to 1 million output tokens per response — that is approximately 750,000 words, depending on how it is calculated.
Bloomberg reported on internal doubts
Bloomberg, citing anonymous sources familiar with the internal assessment, reported that some Google employees consider the model’s practical performance on specific programming tasks to be weaker than its benchmark results suggest. The agency also reported on comments regarding front-end design.
In its response to Bloomberg, Google stated that it would be inaccurate to describe the Gemini 4 Argon as weak in coding. It is, therefore, a matter of disputed assessments rather than an independently verified technical flaw in the model.
Google has decided not to release the Gemini 3.5 Pro following the missed June deadline. In July, the company confirmed that Gemini 3.5 Pro was undergoing testing, but did not officially announce the cancellation of the release. The practical capabilities of Gemini 4 Argon can be assessed more fully once access is expanded, although Google has not yet specified a timeline for this.
