Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

A QA user thanks Mukesh for a complimentary 18-month Google AI Pro subscription worth about ₹35,100, which gave free access to Gemini 4 Argon.

The author reran the same prompt on Gemini 3.8 Flash and Gemini 4 Argon and posted comparison images of the two physics simulation results.

A Japanese user says Argon with deep research produced project ideas several levels above Astra, Opus 5.5 and Grok, but the AI Pro plan hit its limit after two deep research runs.

A Vietnamese post says Gemini 4 Argon is rolling out globally in the Gemini app and Search AI Mode, is Ultra-only on Antigravity, and costs $2 per million input tokens and $10 per million output tokens via API.

The author used Gemini 4 Argon in the Gemini app with a self-built Taiwanese-language skill to generate a dynamic all-Taiwanese webpage explaining LLM principles and workflows, and shared the iteration process for fixing wording.

A video roundup shows Argon turning a hand-drawn sketch into a 76-DoF humanoid simulation, training a control model on 4,500+ simulated episodes to raise real-world success from 56% to 75%, and scoring 87% on the ERQA physical reasoning benchmark.

Google opened Gemini 4 Argon to everyone on October 13, with introductory API pricing of $2 per million input tokens and $10 per million output tokens that doubles after November 30; Google says it ranks first on the Vals index.

Signing in to v0.app with a Google account gives monthly free credits to use Gemini 4 Argon, with no payment or card required.

After Gemini 4 Argon's release, some Antigravity users hit region restrictions; they can check their account's associated country on Google's terms page and request a correction via the country association form.

A post claims Gemini 4 Argon is currently free to use in v0.app, alongside models like GPT-6 Astra and Opus 5.5, with new accounts getting $5 in credits.

The author says 60 lines of local code in agy bypass the Pro limit, using each new Google account's $300 credit to call gemini-4-argon, with switching accounts to keep going.

Citing Apollo charts, the post says Argon's no-CoT implicit reasoning scores NCRI 150.3, second only to GPT-6 Astra, while showing 96.9% information suppression and a 53.9% escape rate from dual monitoring, earning it the label "dirty looper."