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

The author plots Gemini 4 Argon on the AutoResearchExam Pareto curve, reporting near-frontier ML engineering performance at 2–3x lower cost.

Using a colorfigures skill, Gemini 4 Argon built six machines from scratch, each a single HTML file with real 3D-to-SVG projection, flat banded shading and clickable multi-step animations, all one-shotted.

A per-use-case AI recommendation list names Gemini 4 Argon for multimodal understanding, alongside Claude Opus 5.5 for coding and GPT Image 2.5 for image generation.

Sundar Pichai and Kantrowitz discuss Gemini 4 Argon, noting 3x Gemini 3's token volume in its first 72 hours, plus compute trade-offs and the shift to agentic coding.

Google says Gemini 4 Argon handled nearly 25 trillion tokens in its first 72 hours, about 3x Gemini 3's same-period volume.

A paying Google Fi and Google TV user complains that despite Argon being a great model, there is no clear way to get it inside Antigravity.

On stream, Argon used its million-token context to plan a robot project unaided and delegated coding to Opus 5.5 at a lower thinking level; the author calls it the best orchestrator.

The post claims Gemini 4 Argon is the first model to lose LMArena votes weekly, dropping from 4932 to 4900 between Sep 29 and Oct 20, and alleges rankings follow Polymarket bets rather than votes.

Ramp Labs reports Gemini 4 Argon scores 61.1% on Ramp Accounting Bench, 10.6 points above Opus 5.5, with its biggest lead on the hardest tasks.

Sundar Pichai says Argon used almost 25 trillion tokens in its first 72 hours, about 3x Gemini 3.0, that Google is capacity constrained serving it, and that some areas still need improvement.

In a Big Technology Podcast interview, Pichai said Gemini 4 Argon handled nearly 25 trillion tokens in its first 72 hours, about 3x Gemini 3.0, while still compute-constrained, and rejected dark-fiber comparisons.

A reviewer had frontier models critique a full architecture plan; Opus 5.5 scored 95%, Astra 77% and Gemini 4 Argon 74%, testing reasoning, research and hallucination resistance.