Gemini 4 Argon is Google DeepMind's new flagship model, and the clearest way to understand it is by where it wins and where it does not. The model card, which a Two Minute Papers video walked through, describes a large sparse mixture-of-experts model, built from scratch rather than fine-tuned from an older Gemini, that reads text, images, audio and video and writes text, with up to 1M tokens of combined context. Google aims it at software engineering, legal and financial knowledge work, and cyber defence.
The legal result is the standout. On Harvey's Legal Agent Benchmark, Argon scores 19.6%, against 3.8 to 6.7% for GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. The absolute number is low, which says how hard the test is; every model fails most of it. Long documents are another strength, with Argon near the top on long-context tests such as GraphWalks. Agentic coding is the weak spot: on Terminal-Bench 4.0, which tests agents working at the command line, Argon reaches 57.4% against 66.4% for Opus 5.5.
Our review of Argon adds context the launch material does not. Independent scoring by Artificial Analysis puts it level with GPT-6 Astra and behind Opus 5.5 overall, notes a low hallucination rate, and measures a slow start of about 45.6 seconds to the first token. API pricing is an introductory $2 per million input tokens and $10 per million output until 30 November 2026, then $4 and $20. For buyers choosing between frontier models, the pattern suggests Argon for long contracts, filings and research reports, and Claude or GPT for terminal-driven coding agents.
The part deserving more attention than the benchmark table is the safety section. Under its Frontier Safety Framework, Google says it cannot rule out that Argon has reached the critical capability level for biological information risk or for cyber offence, and that it built safety cases arguing the remaining risk is below those levels. It says the same about a threshold for stealth and situational awareness, which concerns a model's ability to evade oversight. These are Google's own assessments of its own model.
The usual cautions apply. Most of the launch figures were run at Argon's highest thinking setting, while the standard model starts lower, and the comparisons are Google's choice of tests. Users in the video's comments describe a model that fills its context window fast and can be hard to steer, anecdotes rather than evidence. The legal lead is large enough to matter; whether it survives independent legal evaluations is the question to watch.