Google DeepMind released Gemini 4 Argon. It is a new frontier model in the Gemini series. The model supports text, images, audio and video inputs. It offers a 1M token input and output context window. It uses a sparse mixture-of-experts transformer architecture. It was trained on Google TPUs using JAX and ML Pathways. It is available via the Gemini App, API, AI Studio, and business channels. It leads benchmarks in knowledge work, agentic coding, science, math and long context. It targets complex workflows in software engineering, legal, finance and cybersecurity. It reduces unjustified refusals and improves refusal tone over prior models. It has a knowledge cutoff of March 2026. It is subject to Google’s Generative AI Prohibited Use Policy.
What changed
Gemini 3.1 Pro was the prior frontier model in the series.
What it unlocks
Running long, multi-step reasoning tasks across text, code, images, audio and video in one workflow.
- 1M token input + output context window
- 68.9% on Vals Index
- 77.9% on DeepSWE v1.1
- 99.7% on GraphWalks up to 128k
Sources