The engineering numbers Google attached to Gemini 4 Argon

Google's October 13 introduction of Gemini 4 Argon attaches a list of measured claims to the launch. Koray Kavukcuoglu's post is the source. Littlewolf did not repeat the runs.

What Google reports

• Vals Index: Argon at number one, as Google states it.
• Antigravity: agentic code submissions up 35%, and rollbacks down nearly a third.
• A Chrome case: a folder of about 6 million lines inside a Chromium tree of about 40 million lines, described as months of work done in weeks.
• ADK: in-memory session prep down 40%, cold imports up 21%.
• xTB: up to 2.3x faster molecular structures and 1.8x on conformer search.
• Gemma 4 E2B: 2.85x inference on an A100 through vLLM, from about 37,900 tokens per second to over 108,000.
• ERQA: 87%.
• Gray Swan: indirect prompt-injection attack success of 0.7% across 15 adaptive attempts.
• Shortcut attempts: 2%.
• A bio red team: unlocked under 1% of bio sub-tasks.
• Adversarial violations reduced nearly twofold.

Google also says 10,000 scientists were offered the highest Antigravity tier free for six months, subject to eligibility, capacity, and selection.

How to read the list

Each line is a vendor result from that post. A higher token rate on one A100 setup, or a lower attack-success rate on one eval, is not a general ranking Littlewolf is willing to sign. Token prices for the same model are in the Argon price note and on AI Model & Plan Prices.