Gemini 4 Argon

Argon keeps a plan in view across a long run. The 1M token context is shared between input and output, so a job can read a large body of material and write a long answer inside the same budget. The continuation protocol lets the model pick up after hitting the response limit instead of starting over.

Inside Google Antigravity, Argon plans changes, edits files, and runs commands. Google reports that teams using it shipped more than 35% more agentic code submissions and saw rollbacks fall by nearly a third. Those numbers come from Google's own engineering groups, so your results depend on your repo and review process.

The Chrome team used Argon to separate core page handling from a legacy folder of about 6 million lines. The refactor touched hundreds of files and took weeks instead of months, according to Google. It also removed a class of browser crashes tied to moving live pages between tabs and windows.

Input covers text, images, video, audio, and PDFs; output is text. That mix matters for contracts with scanned signatures, lab reports with charts, or recorded meetings that need a written summary. Google says the model handles deep document comprehension and longitudinal data analysis with strict grounding.

Function calling, code execution, file search, search grounding, and structured JSON output all work through the Gemini API, along with a preview of computer use. Developers pick low, medium, or high thinking effort per request, which keeps simple calls cheaper.

Argon was trained to pause before irreversible steps and to report negative findings. Google measured shortcut attempts on coding benchmarks at 2%, and the model leads the Gray Swan indirect prompt injection test with a 0.7% attack success rate across 15 adaptive attempts.