Google announces Gemini 4 Argon with a 1 million token output window

Google on Wednesday posted Gemini 4 Argon benchmark and pricing numbers for its next frontier model, the company's most capable general AI system, including a 1 million token output window, how much the model can write and think in one go, up from a prior 64K ceiling, and introductory API rates of $2 and $10 per million input and output tokens, Koray Kavukcuoglu wrote for Google. Cached input tokens are 95% off the input price.

Argon is built for complex, long-horizon workflows across real-world software engineering, enterprise knowledge work in legal and finance, and cybersecurity defense. Google says it is actively engaged in the U.S. government's voluntary process for pre-release model access while gradually expanding who can use the system.

DeepSWE v1.1 scores 77.9%, SOTA for long-horizon software engineering. 9to5Google prints that mark against Claude Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%, and says after the introductory period the price moves to $4 and $20 per million input and output tokens. AutomationBench hits 51.3% and leads the board. LVBench hits 91.7% for long video. CWE-bench v1 hits 68% and ties for first on vulnerability remediation. Google also says Argon leads on the Vals Index.

The first rollout lane is the Fairwind Program, Google's vetted early-access path for trusted cyber defenders. For those defenders and internal teams, Google is releasing Argon without cyber guardrails so they can use full frontier cyber-defense capabilities. Wiz is already using Argon via Scan for Good; an early demo uncovered a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide that previous frontier models had missed. Broader release is framed for paid API customers and Google AI Ultra subscribers after more testing.

Axios quotes Tulsee Doshi, head of Gemini products, calling Argon a well-rounded model with frontier capabilities across several domains. Axios says Argon outperforms GPT-6 Astra on several coding and knowledge-work marks after roughly a year since Gemini 3, and notes that an expected Gemini 3.5 Pro never shipped.

Google's internal color includes a quantum subroutine that beat a published baseline by 40% in minutes, Argon agents that freed more than 300 TiB of memory with an estimated 500 TiB to 1 PiB total, C and C++ to Rust migrations including more than 800,000 lines for the Fuchsia Zircon kernel, and a libgav1 Rust port that ran 2.7× faster after agent work.

After a year of Flash patches while rivals set the flagship pace, the consumer chat box is still waiting on the later wave.

The public scoreboard is already out while the first seats stay invite-only.

Sources

• Gemini 4 Argon – Google
• Google announces Gemini 4 – Axios
• Gemini 4 Argon announcement – 9to5Google
• Google Gemini 4 Argon – The Verge
• Google announces Gemini 4 Argon AI model, but you can't use it yet – Ars Technica