Google has made a move in the artificial intelligence arena. The company has unveiled Gemini 4 Argon , its new cutting-edge model with which it aims to once again compete with the most powerful systems currently available. The announcement comes after several months in which the company seemed to have lost some ground to the advances of OpenAI and Anthropic.
The model is designed for complex professional tasks , such as programming, cybersecurity, or enterprise knowledge work. According to data shared by Google, Argon matches or even surpasses competing benchmarks in several tests, although opinions within and outside the company are not entirely unanimous.
A rival for the GPT-6 Astra and Claude Opus 5.5
The launch of Gemini 4 Argon has a clear objective: to compete with OpenAI and Anthropic . Google's benchmarks place the model ahead of Claude Opus 5.5 and Fable 5.1 in various tests. In the Artificial Analysis Intelligence Index, an independent evaluator, Argon achieves 53 points, the same as GPT-6 Astra, OpenAI's premium model.
In the DeepSWE test, which focuses on real-world software engineering, Argon achieved a score of 77,9% , compared to 74,1% for GPT-6 Astra and 74,2% for Claude Opus 5.5. Google also highlighted its first-place ranking in the Vals index, which measures economic impact in areas such as finance, law, and taxation. And in the CWE-bench, an assessment of vulnerability remediation, it scored 68%, tying for first place.
One of the most striking findings is the reduction in hallucinations . According to Artificial Analysis, the Gemini 4 Argon rate is 15% in AA-Omniscience, while the GPT-6 Astra rate jumps to 50%. The difference lies in the fact that Argon is more likely to answer "I don't know" when unsure, rather than fabricating an answer. However, Astra achieves greater accuracy in the same test, so this data needs to be put into context.
Technical capabilities: one million tokens and autonomous cyber defense
One of the most significant new features of Gemini 4 Argon is its ability to generate up to one million tokens in a single response . To put this in perspective, a token is equivalent to a fragment of a word, and one million tokens is roughly 750.000 words in English. This represents a more than fifteen-fold increase over the previous limit of 64.000 tokens. With such a large capacity, the model can handle tasks that previously required splitting into dozens of requests, such as migrating an entire codebase.
Cybersecurity is another area where Google has focused its efforts. Argon is capable of locating, validating, and repairing flaws in more than twenty programming languages, and can even analyze live web systems without needing access to their source code. The company claims it can find critical vulnerabilities and fix them autonomously, similar to what the Gemini 3.8 Flash Cyber model aimed to do to strengthen programming , although the specific details of its findings have not been made public beyond what Google has chosen to share.
Within Google, thousands of employees already use Argon daily. The company states that its teams use it for research and programming tasks, and that it has freed up more than 330 terabytes of memory in its data centers. These are practical examples the company showcases to demonstrate that it's not all just abstract figures, although the real-world user experience remains to be seen.
Limited release, pricing, and doubts about actual performance
Despite the media hype, the truth is that almost no one will be able to try Gemini 4 Argon yet . Google has opted for a phased rollout, starting with a small group of cybersecurity experts selected through the Fairwind program. The model is also being evaluated by the US government, a standard procedure before launching these types of systems. Later, it will be available to paying API customers who are already Google AI Ultra subscribers, although there is no specific date for its general release.
The starting price is another of Google's arguments to attract companies. During the promotional period, it will cost $2 per million tokens to enter and $10 per million to exit (approximately €1,75 and €8,75). Afterward, the fee will double to $4 and $20. According to Artificial Analysis, with this initial discount, the cost per task would be 40% lower than that of GPT-6 Astra.
It's not all good news. Bloomberg has spoken with Google employees who claim that, while Argon shines in benchmarks, it doesn't perform as well as expected in real-world programming tasks . The company denies these claims and maintains that its teams use it extensively with good results. This discrepancy reflects an uncomfortable truth: synthetic benchmarks don't always equate to day-to-day work, where repositories are massive, dependencies are imperfect, and documentation is scarce.
The political context has also been relevant. A day before the announcement, Google CEO Sundar Pichai, along with US President Donald Trump and other tech executives, signed a voluntary self-regulatory agreement. The document outlines audits and internal controls for cybersecurity and biological threats. Google insists that Argon has border safeguards and will not be released to users until it offers maximum security.
The arrival of this model comes after the promised Gemini 3.5 Pro never saw the light of day. That project suffered from training errors and significantly underperformed compared to the competition, leading the company to abandon it and focus on Argon, which has been under development since July. With this move, Google aims to demonstrate that it remains a contender in the race for cutting-edge AI, although skepticism about the gap between benchmarks and real-world experience continues to linger in the industry.
