Intel / article

Gemini 4 Argon: our next era of frontier intelligence

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

From the Daily Brief

Google DeepMind's Gemini 4 Argon is pitched at long-horizon work: multi-step software engineering, legal and finance tasks, and autonomous patching of vulnerabilities. The headline figure is a 1M-token limit, which DeepMind's own post on X describes as an output limit, enough for a model to finish a long task in one generation instead of stitching together many calls. The catch is access. The first rollout goes to trusted cyber defenders through the Fairwind Program, and Google says broader availability depends on further safety testing, with no date given. For most teams that means Argon is a capability marker rather than something to put in a roadmap. If you run long coding or security workflows, the practical step is to keep your current model choices and harness design portable, so that a switch is cheap when access opens. Treat any benchmark claims as vendor-reported until outside evaluators can run the model.

Read the 2026-10-01 Brief →

Context

Google DeepMind presents a concrete example of Gemini 4 Argon doing iterative engineering work: improving an existing Rust version of its video decoder. According to the announcement, agents repeatedly tested changes and examined compiler output, producing a decoder that ran 2.7 times faster than that Rust baseline while preserving identical video output.

The reported gain concerns that particular optimization task. It illustrates how an can use measurements to revise code through successive experiments. The September 30 announcement places initial access with trusted cyber defenders through the Fairwind Program, with broader availability still planned.

Key takeaways · AI-distilled

Google lists Argon's introductory price as $2 per million input and $10 per million output, with cached input 95% off. Access starts with trusted cyber defenders; paid API customers and AI Ultra subscribers come next.
Google raised Argon's output limit to 1M tokens from 64K, so a single trajectory can generate hundreds of thousands of tokens. It reports 77.9% on DeepSWE v1.1, 51.3% on Zapier's AutomationBench and 91.7% on LVBench.
Google says Argon agents replaced 32K lines of SIMD code in a Rust port of the libgav1 video decoder with safe Rust the compiler auto-vectorizes, making it 2.7x faster with identical output. Other agents freed 300+ TiB of fleet memory.
Trusted defenders get Argon without cyber . Google says it ties for first on CWE-bench v1 at 68%, and that Wiz used it to find a critical data-exposure flaw in hospital software that earlier frontier models had missed.
For safety, Google monitors Argon's and actions to halt out-of-bounds behavior, and says it kept training-run monitor findings out of training so the model would not learn to evade monitoring.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
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 deepmind.google