Gemini 4 Argon targets long-horizon coding

Argon looks like a serious catch-up model, but not an undisputed leader. Its strongest story is sustained agentic engineering—not magic 3D generation—and independent comparisons still show rivals winning key coding and computer-use tasks.

• –Google reports Argon leads its published benchmark table across most evaluated domains, including coding, legal, finance, and science
• –Long decoding and multi-step reasoning make it suited to large refactors, research workflows, and autonomous agents
• –Prompt-generated 3D worlds and games are better understood as agentic software-building demos than evidence of a dedicated 3D simulator
• –Independent evaluations place Argon roughly alongside GPT-6 Astra, with Claude and OpenAI models still stronger in some coding and computer-use scenarios
• –Introductory pricing is aggressive, but Google plans to raise it to $4 per million input tokens and $20 per million output tokens after November 30, 2026