Google launches Gemini 4 Argon for paying users, tying OpenAI’s flagship models

Google is rolling out Gemini 4 Argon today, opening it to developers, enterprises, and paying users across its main products. The model is available through the API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and for Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search.

Google is selling Argon as its strongest model yet for “deep work” on messy, multi-step tasks. The company says it beats other frontier models across benchmarks in areas like finance, coding, legal work, tax work, cybersecurity, and science, and it places No. 1 on the Vals Index. That matters because Google is not positioning this as a chatty demo model. It wants Argon to be the one that does the boring, expensive, error-prone work.

The price is aggressive too: $2 per million input tokens and $10 per million output tokens at launch, with cached input tokens 95% off. For teams that run models all day, those numbers are part of the product, not a footnote.

Inside Google, the model has already been used on real engineering work. The company says Argon in Antigravity boosted agentic code submissions by more than 35% and cut code rollbacks by nearly a third. In Chrome, it helped separate core webpage handling from a roughly 6 million-line legacy feature folder inside Chromium’s 40 million-line codebase, turning months of refactoring into weeks. On Google’s open-source Agent Development Kit, Argon reportedly cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%.

The same pitch extends far beyond software. Google says researchers used Argon to make xTB up to 2.3x faster for molecular structure prediction and 1.8x faster for conformer search, and that it helped improve the speed of open Gemma inference by 2.85x on an A100 with no quality drop. It also cites work in cyber defense, fusion research, rare-disease analysis, enzyme design, and robotics, including an internal test where a humanoid robot sank seven straight cornhole shots after watching a video and having Argon build the simulation code. Google is also giving 10,000 scientists six months of the highest usage tier in Google Antigravity for free, subject to eligibility and selection.

Google pairs all of that with a long safety section, arguing that a more capable model needs heavier guardrails. It says Argon uses a multi-layer defense stack, stronger jailbreak resistance, and real-time monitoring for agent misbehavior. The company is clearly trying to sell this as both a leap in capability and a proof that big models can be made safer at the same time.