Gemini 4 Argon: Our Next Era of Frontier Intelligence

Google has introduced Gemini 4 Argon, a frontier model aimed at the kind of work that doesn’t fit neatly into a single prompt. It is rolling out first to a set of trusted cyber defenders through the company’s Fairwind Program, with broader access coming later after more testing and guardrail work.

Argon is pitched as a model for deep, long-horizon reasoning. Google says it’s already being used internally for specialized coding, research, writing, and other workflows, and that thousands of employees have flagged it as useful in day-to-day engineering work. The company is also putting a price on the thing: $2 per million input tokens and $10 per million output tokens at launch, with cached input tokens discounted by 95%.

The scale is hard to miss. Google is raising the output token limit to 1 million tokens, up from 64,000. That extra room is meant to let the model work through large problems in one run instead of constantly hitting a ceiling and restarting the conversation. On its own benchmarks, Google says Argon reaches 77.9% on DeepSWE v1.1, which measures long-horizon software engineering tasks, and leads the Vals Index, the company’s broader test for work across finance, coding, legal, and tax domains.

Inside Google, the company says Argon has helped with quantum algorithmic optimization, memory tuning across data centers, and code migrations from C and C++ to Rust. In one example, it improved on a published quantum baseline by 40% in minutes. In another, agent analysis helped free up more than 300 TiB of memory, with estimated total savings of 500 TiB to 1 PiB. And for libgav1, Google says Argon replaced 32K lines of SIMD code, producing a memory-safe video decoder that runs 2.7x faster than the Rust port while keeping identical output.

Cybersecurity is the other big push. Google says Argon can autonomously find, validate, and patch vulnerabilities, and that it will be released without cyber guardrails to trusted defenders and internal teams so they can use the full defensive version. The model tied for first on CWE-bench v1 with a 68% score, and Google says Wiz is already using it for its Scan for Good initiative. Before a wider release, Google says it is still strengthening safeguards around misuse, prompt injection, and misalignment.