Google DeepMind showed Gemini 4 Argon — the next generation of its frontier model. The company frames it not as a chat to “hang out”, but as a partner for long jobs: refactoring millions of lines, speeding scientific packages, reading patents and drawings.
What they already did inside
On Chromium, Argon helped split web-page handling from a folder of about six million legacy lines inside a forty-million-line codebase. For the Chrome team that cut months of architecture work to weeks and should let live tabs move between panes without a reload. In the open quantum-chemistry package xTB the same runs gave up to 2.3× faster structure prediction and 1.8× conformer search; patches went into the official repo.
On the ADK agent framework a night run cut preparation of long in-memory sessions by 40% and cold import by 21%. On Gemma inference through vLLM the team claimed a 2.85× speedup on A100 with full token match. Separately Argon moved more than 700 Google Skills teaching labs off hard-wired images — hundreds of engineering hours saved.
Science and defence
The model is sat next to Isomorphic Labs researchers: parsing dense PDFs and patents that used to take weeks compresses to hours. On cyber defence Google talks about first place as a pair on CWE-Bench (68%) for closing vulnerabilities and greater resistance to indirect prompt injection. Those are vendor claims, not an independent audit.
How to look at it
This release is about “long-thinking” code and science, not a new button on a phone. A separate Google post is about robots: the same Argon family learns from video to assemble a physically honest simulator. Watch whether those speedups show up in the public API and what tokens will cost, not the slogan “a new era of intelligence”.
