Google presented its new AI flagship model Gemini 4 Argon on September 30, 2026, aiming to catch up with OpenAI and Anthropic. The system reportedly delivers up to one million output tokens per response and achieves top scores on several programming and security tests. Access is currently limited to selected cybersecurity partners through the Fairwind program.
Argon achieves top scores in programming and security tests
According to Google, Argon is positioned for long-running programming tasks, legal research, and cybersecurity analysis. The model reportedly delivers up to one million output tokens per response – a multiple of the previous limit of 64,000 tokens. In realistic programming tests like DeepSWE, Argon reportedly achieves 77.9 percent, while in detecting known security vulnerabilities in the CWE-bench test, it reportedly ties for first place with 68 percent. The company also cites top scores in video understanding and in defending against prompt injection attacks.
One example cited by Google: security firm Wiz used Argon through its Scan for Good initiative to find a critical vulnerability in widely used clinic software that previous flagship models had missed. In direct comparison with OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5, Argon lags behind on two of four programming metrics, according to Reuters, even though a Google spokesperson describes the model overall as comparable to the competition’s flagship models.
Access remains limited to cyber defenders for now
Argon is not publicly accessible for now. Google is restricting access to selected trusted cyber defenders in its own Fairwind program and is also taking part in the US government’s voluntary process for pre-release model access. Google had already granted the specialized model Gemini 3.8 Flash Cyber exclusively to authorities and infrastructure operators through the same program in September 2026.
For API usage, Google cites an introductory price of two dollars per million input tokens and ten dollars per million output tokens; cached inputs cost up to 95 percent less. After the introductory phase, prices are expected to rise to four and twenty dollars respectively, aligning with Anthropic’s Claude Opus 5.5. The company says it will expand access to paying API customers and Google AI Ultra subscribers “as soon as possible,” without naming a firm date. Google has not yet commented on a possible launch in Germany or the EU.
Google cancels Gemini 3.5 Pro and reshuffles its AI leadership
With Argon, Google is also definitively abandoning the Gemini 3.5 Pro version originally announced by CEO Sundar Pichai for June 2026, which had been repeatedly postponed due to weak internal programming test results. The realignment coincides with a leadership change at DeepMind: co-founder Demis Hassabis handed operational leadership to Koray Kavukcuoglu back in August, and Kavukcuoglu has now authored the Argon announcement himself, saying he is focused solely on staying at the technology frontier.
Internal doubts stand in contrast to the public rollout: according to a Bloomberg report cited by 9to5Google, the benchmark scores do not fully hold up in the daily work of some Google teams, with Argon reportedly still struggling with certain programming tasks. Externally, Google is shifting its message accordingly, from pure top performance toward cost advantages over the competition.
What will matter is whether Google closes the gap between benchmark results and everyday use within its own teams before Argon becomes more widely available. For companies outside the cybersecurity sector, the model remains only a promise for now – Google has not yet scheduled a concrete opening for paying customers or private users in Europe.
