Intel / blog

Gemini 4 Argon: our next era of frontier intelligence

Source: blog.google
Date: Published Sep 30, 2026

Context

Google is positioning Gemini 4 Argon for engineering tasks that require many rounds of analysis and testing. One concrete example is its work on an existing Rust version of the libgav1 video decoder: Argon agents ran many rounds of profile-guided experiments and studied compiler output while replacing 32,000 lines of specialized code. Google reports that the resulting decoder ran 2.7 times faster than that Rust starting point, with identical video output. The comparison is with the earlier Rust port, not the optimized C++ version. Google's account.

The September 30 announcement described initial access for trusted cybersecurity defenders through the Fairwind Program, with broader availability planned after further safeguards work. The reported engineering example therefore illustrates an internal application of Argon, rather than establishing general availability or the same speedup for other projects. Release scope.

Key takeaways · AI-distilled

Google says Argon's output limit rises to 1M from the previous 64K, so a single trajectory can generate hundreds of thousands of tokens of reasoning. The headline 1M figure is output headroom, not only a larger input window.
Argon will launch at an introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% off. Google says access will widen first to paid API customers and Google AI Ultra subscribers after the defender cohort.
In one internal example, Google says Argon agents replaced 32K lines of SIMD code in libgav1's Rust port with safe Rust the compiler auto-vectorizes, producing a decoder 2.7x faster than the earlier Rust port with identical video output.
Google reports 77.9% on DeepSWE v1.1, 51.3% on Zapier's AutomationBench, 91.7% on LVBench and a tied-first 68% on CWE-bench v1. All are vendor-reported figures from the launch post.
Trusted defenders get Argon without cyber . Google says it monitors and actions for misalignment, and avoids feeding monitor findings back into training so the model's reasoning does not learn to evade monitoring.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.

Read the source blog.google