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

The poster shares supposed leak data claiming Gemini 4 Pro beats GPT-6 Astra on most benchmarks, but rates such leaks as unreliable as the earlier Gemini 3.5 Pro leaks.

The poster tries what is rumored to be Gemini 4 Pro disguised as Gemini 3.8 Flash High, but the model calls itself Gemini 3.7 Flash throughout the chat, and jokes about being downgraded.

A roundup of rumors about next-gen models from six companies including Anthropic, Google, OpenAI and xAI, noting Google has only confirmed Gemini 4 pretraining started, not an October Pro release.

The author says MiMo's RL training has finished, with the pro model rising from 58.41 to 72.57 on DeepSWE, still under the top score of 74 held by Astra, Gemini 3.8 Flash and Opus 5.

A reply claims both Gemini 4 Pro and Flash will launch next month, an unconfirmed release-date rumor.

The author gave the same prompt to Gemini 4 Pro and GPT-6 Astra to build an interactive 3D mechanical butterfly in Three.js, and the two results differed completely in UI, structure and rendering.

A list of frontier model launches for next week says Gemini 4 Pro is already in stealth testing, alongside Claude, Grok and GPT-6-Sol releases.

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.