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

Gemini 4 Argon scores 0.4943 on the full 88-task RSI-Exam and 0.5153 on the public 35, ranking third and marking a big jump over Flash models.

A new 'Gemini 3.8-flash-router' slug appeared in Antigravity ACP, speculated to be Gemini 4 Flash being routed, though nothing is confirmed.

Gemini 4 Flash has been spotted inside Antigravity with a 200K context window, suggesting an imminent release.

A user says Gemini 4 Argon raised their AI experience and is using it to build a Python pipeline combining 100 Abhirami Andhadhi songs with commentaries and YouTube discourses.

The author shares a set of instructions that can be added to AGENTS.md to reduce sycophantic responses from Gemini 4 Argon, linking to the original thread.

A side-by-side video shows outputs from Claude Opus 5.5 and Gemini 4 Carbon on the same task.

A user notes Google routes Pro, Ultra and promo Pro to different models and lists internal Gemini 4 Argon and Carbon variants seen in routing.

The post strings together clues: Gemini 4 Flash reportedly in quiet testing and close to Argon, a gemini-3.8-flash-router reference in Antigravity, and Carbon dropping off Arena for one tester, guessing Flash and Carbon land together.

A user reviews the Gemini 4 Flash and Gemini 4 Carbon rumor cycle, arguing Google's tease-to-launch process takes at least half a month, so a pre-Double-11 release would already be fast.

Antigravity docs split AI Pro into trial and paid tiers; students on trial lacked Gemini 4 Argon and other models, and upgrading triggered refunds that wiped their plans.

A user had the unreleased Gemini 4 Carbon generate an animated SVG of Jensen Huang in his signature black leather jacket, joking that even Google's model knows the meme.

A leak claims Gemini 4 Flash is the cheaper Argon variant while Gemini 4 Carbon is 10x better than Argon and Opus 5.5-level.