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

Shahbaztwt shared a screenshot of a Gemini 4 Pro benchmark, saying the first bar is not playing fair, without specifying which test it is.

A short reply to Google AI Studio says the user is waiting for Gemini 4 to be released before starting to build anything.

Quoting Google employee Tulsee Doshi: hype for Gemini 4 is building internally as teams start using the model.

A video says early reports show Gemini 4 Pro pushing hard on reasoning, coding, multimodal understanding and agentic tasks, which could reshape the top of the model race if it holds up in real-world testing.

One user speculates that some Arena results labeled gemini-3.7-flash are actually Gemini 4 Pro, but not every request routes there, which would explain why reported results swing between frontier-level and far behind.

A Japanese-language post says that after Gemini 3.8 went live, Gemini 4 Pro already seems available on Arena, and jokes that the holiday week will be busy.

A user says Gemini 4 Pro produced a beautiful mechanical butterfly from a three.js prompt, and notes it had been tested on Arena under the gemini-3.7-flash placeholder before apparently being removed.

A Chinese post says Gemini 4 Pro generated a three.js mechanical butterfly in ten minutes while Fable 5.1 took about half an hour, and claims Google used recursive self-improvement in training.

A single prompt produced a fully procedural, disassemblable two-spool turbofan demo with FADEC and N1/N2/EGT gauges, leading the author to suspect the model is actually the unreleased Gemini 4 Pro.

The poster compares GPT-6 Astra with Gemini 4 Pro and says Astra delivered better results right from the start, without giving specific tests.

LuminaBench says that in Arena comparisons, Gemini 4 Pro outperforms GPT-6 Astra on some specific SVG generation requests and vice versa, and that it is not identical to muse spark 1.3.

The author argues Google doesn't need an Astra-level model to recover: Gemini 4 Pro only needs to match Opus 5 and GPT-5.6 Sol while being cheaper, faster, and offering higher limits to win back market share.