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

A French reply notes that people aren't mentioning that the impressive result came from the new Gemini 4 Pro rather than Gemini 3.1 Pro, hinting 4 Pro is already quietly live.

The author jokes that Gemini 3.5 Pro hasn't even landed while the community is already making up Gemini 4 Pro benchmark numbers, and says to wait for a model card and evals first.

The author says Antigravity's Gemini 3.8 Flash quota has been cut again, a Pro account is no longer enough, and a Gemini 4 Pro launch would likely force an upgrade to Ultra.

The poster says they are still not convinced by Gemini 4 Pro hype but hopes Google releases a real frontier model that can compete with Claude Opus, GPT Sol and Astra.

The poster shares a screenshot of the Artificial Analysis intelligence leaderboard, says Gemini ranks only 14th despite benchmaxing, and asks whether Gemini 4 Pro will lift it much.

The post argues Google chose to focus on everyday tasks rather than raw capability, and says recent leaks point to some kind of RSI breakthrough in Gemini 4.

The poster lists rumored releases including Gemini 4 Pro, Grok 4.7, Opus 5.5, the GPT-6 family and Fable 5.2, and says Opus 5.5 leaks show 20% better token efficiency than its predecessor.

The author argues the $200 subscriptions aren't worth it and hopes Gemini 4 Pro and Flash 4 keep Google DeepMind's pricing advantage.

Users note that Gemini 4 Pro, Grok 4.7, GPT-6 Sol and Opus 5.2 have been rumored for weeks without launching, and speculate the labs are waiting for rivals to go first to benchmark and fine-tune.

A video shows the Arena model assembling a 3D structure of the Airbus H145 helicopter in roughly 10 minutes, though its bottom UI looks rough, and Google has not yet released Gemini 4.

A leaker says Gemini 4.0 is the next major release and will include a Pro model to replace 3.1 Pro, but Google's focus is shifting to Flash, so Pro may only update once a year.

A joke that Gemini 4 Pro is not yet released, while other LLMs already behave as if they know its benchmark scores in advance.