On September 30, Google announced the launch of a new generation of flagship artificial intelligence models Gemini and Argon, which are focused on enterprise knowledge work in software engineering, law, finance, and cybersecurity defense. The models will be made available in phases, with the first group of users being designated cybersecurity partners within the Fairwind program. Subsequently, they will be provided to paid API customers and Google AI Ultra subscription users. The company has not yet announced a specific date for full availability.
For complex, multi-step enterprise tasks
Google positions Argon as a model capable of conducting in-depth reasoning continuously within complex and time-consuming work processes, with applications ranging from code debugging, algorithm design, financial research to drafting legal documents.
To support longer tasks, the output limit for Argon has been increased from the previous 64,000 token to 1 million token. Token is the unit of measurement for the model's processing and generation of information. Google states that this adjustment allows the model to conduct more in-depth reasoning and generate longer content in a single task.
According to the evaluation results published by the company, Argon achieved a score of 77.9% in the DeepSWE v1.1 tests that measure the performance of real, long-term software engineering tasks; in the AutomationBench tests that assess the end-to-end execution capability of core business processes, it scored 51.3%. Google also stated that the model performed leadingly in specialized evaluations such as multi-step financial research, legal research, and drafting.
Already used for internal code migration and data center optimization.
Argon has been incorporated into Google's internal workflow before its external expansion and opening to the public, and is used for daily debugging, large-scale code migration, and infrastructure optimization.
Google revealed that the Argon intelligent agent identifies and implements memory optimizations by analyzing the operational monitoring data of data centers. Since its deployment, it has freed up more than 300 TiB of memory capacity. At the same time, related intelligent agents are migrating some C/C++ code to Rust, which involves core software libraries as well as the Zircon kernel of the Fuchsia operating system, with the latter comprising over 800,000 lines of code.
These major code overwrites still need to undergo automated and manual auditing, simulation testing, and review before they can be used in a production environment.
The new model also continues the previously postponed updates to the flagship products. Google had planned to launch Gemini 3.5 Pro in June of this year, but that version was not released on time; thereafter, the company successively launched smaller Flash series models. This time, they have directly announced Gemini 4 Argon.
The first batch focuses on network security defense.
The Fairwind initiative aims to enable key infrastructure defenders such as governments, medical institutions, and telecommunications service providers to use advanced models in advance to detect and fix vulnerabilities. The program currently has over 650 global partners, some of whom will be granted Argon usage rights and will be able to conduct vulnerability research and repairs through the code security agent CodeMender.
The scope of use for partners is limited; organizations are only allowed to provide access rights to their internal network security, incident response, or penetration testing teams. It is also necessary to record the usage by employees. Access rights to the model cannot be resold or redistributed to external parties.
Google stated that the company is participating in the voluntary access program before the release of models by the U.S. government and will continue to improve protective measures based on feedback from early testers. The measures disclosed include monitoring the model's reasoning process and actions, stopping execution when behavior exceeds user intent, as well as enhancing the isolation protection of the testing environment.
Announcement of promotional period prices; stock price rises after the market closes.
Google announced a promotion period for Argon. During this period, the cost is $2 per million inputs and $10 per million outputs for token. The price for cached inputs is 95% lower than that for regular inputs. After the promotion period ends, the prices for inputs and outputs will increase to $4 and $20 respectively. The company did not specify how long the promotion period would last.
After a limited rollout, Google plans to first expand the service to paid API customers and Google AI Ultra subscription users, before gradually covering more developers, enterprises, and consumers. As of the time of this announcement, the company has not provided a clear timeline for the public version or its launch for cloud computing customers.
Alphabet's stock price rose by more than 1% in after-hours trading on September 30, reaching $349.95. As of this release, the stock has gained approximately 10% for the year so far, but it is still below the high of $408.61 reached on May 18.
