Google has just announced the launch of the next generation of its flagship model, the Gemini 4 Argon.
Google has positioned Argon as a model for complex, long-cycle workflows, focusing on business knowledge such as software engineering, law and finance, and cybersecurity defense.
Argon will adopt a phased release strategy, unlike its previous approach to open to the public directly. The model will be the first to be delivered to trusted cybersecurity defenders through the Fairwind Program.
In terms of pricing, the Argon API is priced at $2 per million of Token inputs and $10 per million of Token inputs, while the caching of Token inputs is 95% lower than the standard input price.
Single-output limit is increased to 1 million tokens
The Gemini 4 Argon's most notable change is to increase the Token output limit from 640,000 to 1 million.
One million tokens refers to a single output ceiling, and the above and below window are two indicators. Higher output space means that the model can continue to reason on a mission trajectory and generate hundreds of thousands or even close to a million tokens. These capabilities are primarily geared towards large-scale code migration, deep research, and multi-step enterprise processes.
Some netizens note that the output limit for most front-end models is about 128,000 tokens. In contrast, the ability of Argon to generate 1 million tokens at a time is rare.
According to Google, Argon scores 77.9% in the DeepSWE v1.1 software engineering rating, which is the current highest. The rating is a major measure of the model's ability to handle real, long-term software engineering tasks.
In terms of business knowledge, Argon ranks first in the Vals Index, which covers areas such as finance, programming, law, and taxation, and is weighted according to the contribution of each industry to U.S. GDP.
Argon has also led ratings such as Vals Finance Agent v2, Harvey Legal Agent Benchmark and Zapier's AutomationBench, with AutomationBench scoring 51.3%.
Argon also emphasizes visual comprehension. Google says that models can analyze professional charts, understand long videos, and act on multiple documents. In the LVBench long video comprehension rating, Argon scored 91.7%.
Entered into Google's internal engineering process
According to Google, Argon has been used by thousands of employees in their daily work, with applications including code modeling, algorithm design and large-scale code library migration.
In the field of quantum computing, Argon helps researchers optimize the time-space resources in quantum algorithms, i.e. multiplying the number of qubits by the number of door operations. Google says that in one test, Argon will improve the published benchmark results by 40% in just a few minutes.
In a data center memory optimization project, Argon analyzed network-wide performance monitoring data and automatically identified and implemented optimization schemes. More than 300 TiB of memory was released after the program went online, with total savings expected to be 500 TiB to 1 PiB.
Argon has also been involved in the migration of Google's internal C/C++ code library to Rust, covering core libraries such as re2 and libgav1, and extending to the Fuchsia Zircon kernel with over 800,000 lines of code. These migrations still require automated auditing, human reviews, and simulation testing due to critical infrastructure.
In the open-source video decoder libgav1 project, Argon agents rewrote about 32,000 lines of SIMD code through multi-round performance testing and compiler analysis. The new Rust version runs 2.7 times faster than the original Rust version when the video output is consistent.
Cybersecurity capabilities in parallel with real-world testing
Cybersecurity is one of Argon's top priorities. According to Google, the model is capable of automatically detecting, verifying and fixing critical software vulnerabilities.
The cybersecurity company Wiz has used Argon for its "Scan for Good" program to screen public infrastructure for high-risk exposures for free. Google said that Argon had discovered a serious vulnerability in medical software used in hospitals around the world that could expose sensitive personal information, a risk that previous front-end models had not identified.
In the CWE-bench v1 vulnerability correction assessment, Argon was ranked first with a score of 68%.
Google also said that Argon performed better than the Gemini 3.8 Flash Cyber in internal vulnerability testing covering 20 programming languages, as well as black box penetration testing that did not provide source code.
As the model improves, Google is also strengthening security mechanisms. Argon will limit cyberattacks and chemical, biological, radioactive and nuclear-related abuse requests, and identify potential risks through in-house activated monitoring, automated Red Team testing and artificial assessment.
Google has also deployed monitoring mechanisms for model mismatches, tracking the model's reasoning process and action trajectory, suspending mission execution when necessary, and isolating high-risk training and assessment environments.
However, there is still a gap between evaluation results and real work experience.
According to Bloomberg, there are still questions within Google about the performance of the Gemini 4 in key tasks such as coding. People familiar with the model said that while it excelled in industry benchmark testing, some coding tasks are still difficult to stably complete when actually used by employees.
If Argon can demonstrate its long-cycle reasoning and complex task execution capabilities, more real-world scenarios are needed.
Google said that the company is participating in a voluntary pre-release process of models promoted by the US government, with the first testers helping Google evaluate model performance and further improve security precautions.
Following the completion of early testing, Argon will be gradually open to developers, businesses and consumers. The first open users will include paid API customers and Google AI Ultra subscribers.
The link is as follows: