Gemini 4 Argon raises Google's output cap from 64K to 1M tokens for single-pass agent jobs

Inside Google, Argon agents are rewriting the Fuchsia OS Zircon kernel from C/C++ into Rust. The kernel runs past 800,000 lines, and the rewrite goes through automated and manual auditing, emulation testing and review before any of it reaches production ^[8].

According to 9to5Google, the new headroom lets Argon generate hundreds of thousands of tokens in a single trajectory and solve tough problems in one go ^[2]. Ars Technica reports Google's version as completing more daunting tasks in a single step ^[15]. The new ceiling is 15.6 times the old one ^[1]. So far the users are thousands of Googlers, by Google's count ^[14], and the largest job Google describes goes to auditors before production ^[8]. In Google's own migration process, the pass is where review starts ^[8].

The output limit also caps what one call can bill. At the list price 9to5Google reported, a response that uses all 1M tokens costs $20 in output, against $1.28 for one that stops at the old 64K limit at the same rate ^[3][4]. At the introductory rate the full response costs $10 ^[2]. Failed runs bill as well. Google says Argon ranks first on Zapier's AutomationBench, which measures end-to-end execution across business functions, with a score of 51.3% ^[10]. An agent that goes wrong 900,000 tokens into a list-price run has spent $18 before anyone reads a line ^[5].

The cyber tiers sort buyers by who they are. The defenders using Argon today came in through Google's Fairwind Program ^[4]. On CWE-bench v1, which tests whether a model can remediate security vulnerabilities, Google says Argon ties for first at 68% ^[11]. Buyers outside the trusted tier get a model Google watches from the inside, with monitoring of internal activations aimed at cyber and CBRN misuse ^[12]. Google's announcement, as reported, does not say how a security team qualifies for Fairwind or what terms come with guardrail-free access.

In my view Argon is worth piloting first on a job that already produces more than 64K tokens of output and already has a human review step. Those teams gain the most from fewer chunks and are already paying for the reviewer. The tradeoff is that each failed run can bill up to $20 of output at list price before the reviewer sees it ^[3].

Two axes sort the decision. One is whether a job needs more than 64K tokens of output in a single response. The other is whether it is defensive security work a cyber guardrail could block. Short output outside security: little changes until a price holds past the introductory period. Long output outside security is the case Google built Argon for ^[1], and the budget line is per run, with $20 as the ceiling. Short security work turns on the tier, so the first conversation with Google is about Fairwind. Long security work needs a written price and Fairwind terms before any pilot result means much.