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

The same voxel pagoda prompt run through Sonnet 5.5, Gemini 4 Pro, GPT-6 Sol and Luna shows clear style differences, with Gemini 4 Pro using checkpoint 2 from its Arena run a few days earlier.

A user says a Gemini 4 Pro that beats GPT-6 Sol and Opus 5.5 would be welcome, but complains that Google's Antigravity 2.0 development environment is so poor that a stronger model alone won't help.

A person who says they spent three years doing pretraining research at Google announced their resignation, saying the company is racing toward self-improving superintelligence and that Gemini 4 will end humanity.

After trying Gemini 4 Pro, Gemini 3.8 Flash, Opus 5.5 and GPT-6 Astra side by side on Arena, the author ranks the first two strongest, places Astra in the middle and finds Flash better than its price suggests.

A short post with no images claims Gemini 4 Pro, Sonnet 5.5 and Grok 4.8 launch today, offering no source or evidence, so it remains a release rumor.

Rumored Gemini 4 Pro benchmark scores look unusually strong, and if even partly accurate would mark a major upgrade for Google, though nothing is officially confirmed.

The author relays rumors that Gemini 4 Pro will launch on October 1 with off-the-charts benchmarks, but says they aren't convinced yet and will wait for the official release.

A long Japanese post lists leaked Gemini 4 Pro scores including 88.7% on DeepSWE, 95.3% on Terminal-Bench, 86.8 on OSWorld and a GDPval-AA lead, plus a 2M context and low price at roughly 60% confidence.

A leaked table shows Gemini 4 Pro beating GPT-6 Astra and next-gen Claude models on DeepSWE, Terminal-Bench and OSWorld, with a 2M context and $2.25/$11.25 pricing.

A user reports that Google AI Plus no longer offers Gemini Pro models without any notice, speculating it makes room for a stronger Gemini 4 Pro reserved for Pro and Ultra members.

A user says the Gemini 3.8 Flash on Arena is actually a Gemini 4 Pro checkpoint that performs decently but refuses large project work.

A leaked benchmark image suggests Gemini 4 Pro would beat GPT-6 Astra and Claude Opus 5.5, but the author doubts it, citing Google's history of benchmark-maxxing and RL on specific skills like SVGs.