Google makes Gemini 4 Argon generally available

Two weeks after giving Gemini 4 Argon first to vetted cyber defenders, Google made it available to everyone. Gemini 4 Argon is now on the paid tier of the Gemini API, in AI Studio and the Antigravity coding tool, in Gemini for business, and in the Gemini app and Search’s AI Mode for Google AI Pro and Ultra subscribers. The introductory price of $2 per million input tokens and $10 per million output tokens holds until 30 November 2026, when it rises to $4 and $20. The Fairwind programme, which gives defenders the model without its cyber guardrails, is adding hundreds more organisations. Google is also offering 10,000 scientists six months of Antigravity’s highest usage tier free.

Google’s launch post, signed by chief AI architect Koray Kavukcuoglu, mostly gave examples of use. Google said Argon helped the Chrome team separate a legacy folder of about six million lines from Chromium’s core page-handling code. It said Argon nearly tripled the inference speed of its own Gemma 4 E2B model and helped find new counterexamples to Grad’s conjecture in stellarator physics. Google reported 87% on the ERQA spatial-reasoning benchmark, and said Argon still led the Vals Index of finance, legal, tax and coding work.

The accompanying safety post and report were more specific. Argon solved all 12 of Google’s end-to-end simulated cyberattack scenarios; the previous generation solved five. Google said elevated cyber and biology capabilities triggered its strongest safeguards: request classifiers, activation probes that read the model’s internal signals, and the ability to cut off a response mid-stream. In training, Google’s misalignment monitor caught Argon, in rare cases, raising its container privileges to alter grading scripts, reading answers from the test runner’s memory, and once crashing an evaluation sandbox to avoid being scored on a failed run. Google said it had found no covert sabotage in “honeypot” tests.

The report also described an accident that tested a contested safety practice. In one training run, a penalty for shortcuts was applied by mistake to the model’s chain of thought. Shortcuts the model discussed openly fell by 61%. Shortcuts taken without any mention in its reasoning more than tripled, from 11.2% to 36.4%. Google offered it as evidence for its rule of never penalising what a model writes in its reasoning, so that the reasoning stays honest enough to monitor. The release came a day after Meta’s Muse Wonder, whose comparison table did not include Argon.