Google releases Gemini 4 Argon, its new flagship for long, complex tasks

Google released Gemini 4 Argon without much warning, positioning it as the new flagship for long, complicated tasks rather than quick chat responses. The headline number is the output limit: Argon can generate up to 1 million tokens in a single run, a sixteen-fold jump from the 64,000-token ceiling on its predecessor. According to Google’s own announcement, that expanded limit is paired with real gains in capability rather than just a bigger buffer, the model posted 77.9% on DeepSWE v1.1 and 91.9% on Vibe Code Bench, both new highs among frontier models, with the Vibe Code result edging out GPT-6 Astra and both current Claude models by a narrow but real margin.

Google isn’t just shipping Argon externally, it’s using it internally first. Argon agents are already working through Google’s own C and C++ codebases, rewriting them in Rust, with one migration reaching the Fuchsia Zircon kernel at more than 800,000 lines of code. That’s the kind of unglamorous, high-stakes engineering work companies usually hand to senior staff rather than an AI agent, and Google is treating the model’s performance there as a credibility marker as much as a product feature. One optimization an Argon-assisted team found freed over 300 TiB of memory across Google’s data centers once it rolled out, with the company estimating 500 TiB to a full petabyte in savings once it’s applied more broadly, a number large enough to matter at hyperscale even before counting the compute cost of running the model itself.

Pricing starts at $2 per million input tokens and $10 per million output tokens, introductory rates that Google has already said will double to $4 and $20 once the launch window closes. That’s a candid way to price a new model, most companies let introductory pricing quietly expire rather than publishing the increase in advance. Access is narrow for now: Argon is rolling out first to cybersecurity researchers through Google’s Fairwind Program, with developers, enterprises, and everyday users told to expect access later rather than immediately. For a model built around long, complex tasks rather than quick answers, gating the rollout to the people most likely to stress-test it with exactly that kind of work is a reasonable way to find problems before a wider release.