Key facts
- Google announced its new flagship AI model, Gemini 4 Argon.
- Argon reportedly used fleet-wide telemetry data to save Google 300 TiB of memory.
- Argon agents migrated C/C++ codebases to Rust across Google, including thousands of lines in core libraries and the Fuchsia OS Zircon kernel.
- Gemini 4 Argon achieved 77.9% on the DeepSWE v1.1 benchmark, surpassing GPT-6 Astra, Fable 5.1, and Opus 5.5.
- Google claims Argon has industry-leading performance in coding, knowledge work, and cybersecurity.
- Gemini 4 Argon will support an output limit of 1 million tokens, up from 64,000 in previous models.
Google announced its new flagship AI model, Gemini 4 Argon, aiming to regain a competitive edge against rivals like Anthropic and OpenAI. The company claims Argon offers industry-leading performance in coding, knowledge work, and cybersecurity, though it is currently only available for internal use and select cybersecurity partners. Google engineers are already utilizing Argon, which reportedly helped save 300 TiB of memory across data centers by analyzing fleet-wide telemetry data. Argon agents have also been employed to migrate C/C++ codebases to Rust within Google, including significant contributions to core libraries and the Fuchsia OS Zircon kernel.
To support its performance claims, Google presented benchmarks showing Gemini 4 Argon achieving 77.9% on the DeepSWE v1.1 software engineering benchmark, surpassing competitors like OpenAI's Astra and Anthropic's Opus. The company also highlighted Argon's leading score on the Vals Index economic analysis test for long-horizon tasks.
While a public release date and API pricing remain unannounced, Google confirmed that Gemini 4 Argon will feature a significantly increased output limit of 1 million tokens, a substantial upgrade from the 64,000 tokens supported by previous Gemini models, enabling users to tackle more complex tasks in a single step.