Abstract:
Google launches a new generation of cutting-edge model Gemini 4 Argon, which is another important model update since the company released Gemini 3 at the end of 2025. Argon targets complex, multi-step tasks, focusing on real-world software engineering, enterprise knowledge work, and network security defense. Google says it can reduce ambiguity, repetition and clichés in responses and compete with industry-leading models in tests including programming, science, finance, law and cybersecurity.
Argon is gradually entering Gemini application and Google search AI mode, first for Google AI Pro and Ultra subscribers; Google AI Ultra users can also use it in the agent development platform Antigravity. The report did not confirm when Plus subscribers or free users will gain access. Google has previously stated that the model will first be open to trusted network security defenders through the Fairwind program, and will gradually expand to developers, enterprises, and consumers after strengthening security protection; therefore, the current opening is still in stages and cannot be considered to be available to all users.
An important specification of Argon is the maximum output word limit of 1 million, which is significantly higher than the previously announced 64,000 word words. Google said this enables the model to continuously analyze and generate a large amount of content in a single task process, and is suitable for long-range programming, legal research, financial analysis and multi-round agency work. It should be noted that 1 million refers to the upper limit of output tokens and is not equivalent to the context window length.
Internal test results released by Google show that Argon scored 77.9% in the DeepSWE v1.1 test that measures long-range software engineering tasks, 51.3% in Zapier's AutomationBench business process test, and 91.7% in the long video understanding test LVBench. In terms of network security, Google said that the model tied for first place in the CWE-bench v1 vulnerability repair test with a score of 68%. The above are all test results disclosed by Google. The test set, task design and evaluation conditions will affect the results and should not be directly regarded as a guarantee of performance in all actual scenarios.
Google also introduced several internal use cases. The company said that the Argon agent has assisted in migrating some C/C++ code libraries to Rust, ranging from tens of thousands of lines of basic libraries to more than 800,000 lines of Fuchsia Zircon kernel; for the video decoder libgav1, the model rewrote about 32,000 lines of SIMD code, and the generated Rust implementation is said to be 2.7 times faster than the original Rust ported version under the same output. The company said the relevant code still needs to go through automated and manual audits, simulation testing and review before being put into production. Google also said that Argon is used to identify and patch security vulnerabilities, optimize data center memory, and assist in debugging the flight physics simulator of its Wing drone delivery company; the model also helps the YouTube comment system cut the required computing resources in half. These effects are internal cases disclosed by the company and are not independent evaluation conclusions.
API pricing for developers is reported to start at $2 per million input words and $10 per million output words, with caching of input words offering a 95% discount on the input price. After the promotion period ends, the standard price will rise to US$4 per million input words and US$20 per million output words. Per-term pricing is not exactly equivalent to the actual task cost, because different models may consume different numbers of tokens to complete the same task.
Because Argon has stronger network security capabilities, Google focuses on strengthening security protection before opening it to the public, including resisting prompt injection, monitoring potential target deviations, and restricting its use in network attacks and chemical, biological, radiological, and nuclear-related malicious activities. Google said it will continue to improve the security mechanism with feedback from trusted testers.
Comments