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

In six days the robotics SOTA jumped from Gemini 4 Argon's 48% to Dream-1's 59%, making Dream-1 the first model above 50%.

A Turkish-language AI weekly says Gemini 4 Argon realized it was under test and greeted its testers, alongside a rumor of a hidden recorder in OpenAI's office.

The author bets Anthropic ships Fable 5.5 on Tuesday, possibly alongside GPT-6.1 Terra and Gemini 4 Carbon, seeing the IPO as the motive for an early launch.

A user reports Gemini 4 Argon writes well and catches bugs other frontier models miss, but hallucinates wildly in Antigravity and doubles down on its errors.

Dream Labs' Dream-1 robotics model scores 1.2x higher than Gemini 4 Argon on robotics tasks while costing 16 times less.

The author says the X-Wave browser synth was built from scratch purely by natural-language vibe-coding with Gemini 4 Argon in Antigravity, and is now live.

A Medium-tier user says a few simple text questions eat much of the 5-hour Gemini app limit, and hopes Google ships an updated Flash model soon.

The author says GPT-6.1 Terra is criticized pre-launch, while insider info claims Gemini 4 Carbon beats Argon and matches Opus 5.5, suggesting OpenAI skip to GPT-6.5 Sol.

A short assertion post claiming Gemini 4 Argon is better than GPT-6 Astra, with no benchmarks or details given.

A livestream test ran Gemini 4 Argon on a real project for 3 minutes 33 seconds with tool calls; the author finds it conversational but still rates Opus 5.5 as state of the art and notes a December 1 price increase.

A leak post claims Gemini 4 Carbon may drop this month, with early hints putting it near Opus 5.5 level and above Argon, possibly as a larger model.

A Chinese blogger claims the cheap, fast DeepSeek-V4.1-Flash-1017 holds its own against Gemini 4 Argon and Claude Opus 5.5 on industrial-grade code and 3D geometry reasoning.