Google announced its Gemini 4 Argon AI model, restricting early access to internal staff and select testing partners, Ars Technica reported. The move follows a summer spent deploying smaller Flash models after the company promised Gemini 3.5 Pro in June.
Engineers inside Google have already deployed Argon to analyze fleet-wide telemetry data, saving 300 TiB of memory across company data centers. Automated Argon agents also migrated C and C++ codebases to Rust, converting thousands of lines in the core re2 and libgav1 libraries and over 800,000 lines in the Fuchsia OS Zircon kernel.
Argon scored 77.9 percent on the DeepSWE v1.1 software engineering benchmark, surpassing GPT-6 Astra, Fable 5.1, and Opus 5.5. The model also led the Vals Index test for economic analysis, which Google cited as evidence of its performance on long-horizon tasks.
Customers accessing the API during a limited promotional window will pay $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95 percent. Output capacity reaches 1 million tokens per task, expanding on the 64,000-token maximum supported by previous Gemini models.
Security partner Wiz is using the model under the Fairwind Program to evaluate cyberdefense capabilities, identifying a critical flaw that exposed personal data across global hospital networks. Google stated that competing frontier models missed the defect, but the company did not provide specifics on those tests. To address model alignment following hacking incidents over the summer, Google integrated monitoring systems that review Argon's chain-of-thought reasoning and halt the model if it acts outside operational parameters.
Subscribers to Google AI Ultra and paid API users will receive access during the first stage of general availability after cybersecurity trials finish. Google will eventually offer Argon to broader enterprise and consumer groups, though it outlined no release dates for those tiers.
