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

A roundup of this week's rumors covers Sonnet 5.5, GPT-6 Cyber, DeepSeek V5 and Qwen 4, with Gemini 4 listed as an October release rumor.

An unsourced prediction post argues that if Gemini 4 really beats GPT-6 Astra and Opus 5.5, Google is back in the frontier top tier.

The author argues that if the Gemini 4 Pro leak holds, frontier-model pricing can't be sustained and a pricing war between labs has already begun.

The poster dismisses the circulating Gemini 4 Pro benchmarks, saying it is good but still behind Opus 5.5 and GPT-6 Astra in many areas.

A short post with no image claims a leaked ranking puts Gemini 4 above Opus 5.5 and Astra 6, without citing a source or data.

A leak claims Gemini 4 Pro is priced at $2.25 per million input tokens and $11.25 per million output tokens, well below GPT-6 Astra and Claude Opus 5.5 if accurate.

A leak claims Gemini 4 Pro scores 88.7 on DeepSWE and beats Astra on Terminal-Bench 2.1, OSWorld 2.0 and GDPval-AA, with a 2M context window and $2.25/$11.25 pricing.

A user says they want Gemini 4 Pro to win but the leaked benchmarks are obviously fake, since the current flagship is nowhere near Opus 5 or GPT 5.6.

A Czech-speaking user says Gemini 4 Pro benchmark leaks are circulating on X, calling them a possible class-jumping comeback if the model doesn't slip to spring 2027.

A poster warns the Gemini 4 Pro leak could turn into another Le Chaton Fat situation, noting the demos look insane but the table looks too clean, and says they're watching with one eye closed.

A poster claims leaked Gemini 4 Pro numbers dominate both Anthropic and OpenAI, joking that rivals now have to price-match and that an IPO is delayed.

In a 3D floatplane physics simulation, Gemini 4 Pro (arena) shows sharper reflections, richer wave detail and stronger lighting than Claude Opus 5.5, whose water looks flat by comparison.