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

A post lists six frontier models that could launch next week, including Gemini 4 Pro, GPT-6 Sol/Luna and Opus 5.5.

The same prompt for an interactive 3D jet engine cutaway produced a far more detailed result from an anonymous Arena model, which the community suspects is Gemini 4 Pro, though Google has not confirmed it.

The author sorts six rumored models by how strong the evidence is, noting Gemini 4 Pro has real Arena checkpoint outputs but still no model card or API ID.

A user reports that Gemini 3.8 Flash on the API shows unusual tps and knowledge cutoff, suspecting it is actually routed to Gemini 4 Pro, and shares video evidence.

A developer asks whether adding Gemini 4 Pro to Antigravity could let it seriously challenge Cursor in the AI coding tool race.

A user says Gemini 4 Pro performs impressively at 3D generation and calls it very tempting, without sharing any specific output.

The author shares a video generated with Gemini 3.1 Pro and says that if the leaked Gemini 4 Pro benchmarks are real, its output is hard to imagine.

The author shares a Three.js scene made with GPT Astra Ultra, saying Astra remains the best pick for Three.js and questioning how strong Gemini 4 Pro really is.

The author ran the Will Smith eating spaghetti test in Three.js, was impressed by the result, and speculated that Gemini 4 Pro's frontend generation is nearing Fable 5.

YouWare built a 3D version of the classic AI stress test, pitting Gemini 4 Pro against Fable 5 Max side by side and calling the results beyond expectations.

A poster lists four models that could ship next week, placing Gemini 4 Pro alongside Opus 5.5, Grok 4.7 and GPT-6 Sol.

An AI daily digest says Sonnet 5.2, Opus 5.2, Fable 5.2 and Gemini 4 Pro are all in testing, alongside news that GPT-6 Astra solved a long-open math problem.