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

A post argues AA Intelligence Index v5 is a bad benchmark because GPT-6.1 Terra outranks Gemini 4 Argon (high) and Claude Fable 5.1 (max).

Google Antigravity IDE now includes Gemini 4 Argon, yet users on the 18-month JIO AI Pro plan still cannot access it in AGY.

The new Artificial Analysis Intelligence Index v5 ranks GPT 6.1 Terra Max above Gemini 4 Argon High, prompting the poster's frustration.

The new AA Intelligence Index v5 puts GPT-6.1 Sol ahead of Gemini 4 Argon and Claude Fable 5.1, which the poster calls suspiciously favorable to OpenAI.

A demo shows Gemini 4 Argon breaking a task into fine-grained text steps fed to Genie 4, which updates the simulation in real time as a humanoid plays checkers.

Artificial Analysis releases v5 results across nine evaluations, with no single leader; Gemini 4 Argon leads AutomationBench-AA at 78% while Claude Opus 5.5 tops three.

An expert-call summary says Gemini 4 Argon was finalized ~1.5 months before launch to compete with Claude 3.5 Sonnet on price/performance, but TPU v8i delays and compute scarcity led Google to suspend it internally and downgrade free users from Flash to Light.

Google says Gemini 4 Argon is gradually rolling out in Antigravity IDE, with eligible Pro and Ultra users getting access within 24 hours.

A user with work access to Gemini Argon Pro says it is powerful but has no middle ground, producing either huge analyses or tweet-length answers.

A long breakdown of Genie 4: live editable worlds, native audio, more realistic humans, with Gemini 4 Argon decomposing tasks and 3.8 Flash directing, still research preview at ~2 minutes of consistency.

Genie 4 adds real-time ambient audio and character dialogue, holds scenes for about two minutes, runs on Gemini 3.8 Flash, and pairs with Gemini 4 Argon for robot training.

A user says engineering students on Google Pro can upload exam papers to Gemini's Pro or Argon mode to get solutions, while NotebookLM struggles with non-theory material.