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

With Opus 5 running at Max effort on the same prompt, Gemini 4 Pro still kept up, which the author sees as a sign it is now competitive for complex visual coding.

A butterfly video generated with Gemini 4 Pro, shown as a demonstration of its video generation output.

Based on the outputs seen so far, the poster says Gemini 4 Pro already matches GPT-6 Astra, and if true, upcoming GPT-6 family models are in trouble.

A tester says Gemini 4 Pro and GPT-6 Astra Pro are basically tied on SVG generation, calling the results the best seen from any model so far, without giving specific examples.

The poster notes that GPT-6 Pro was evaluated in the ChatGPT harness and Gemini 4 Pro in the arena.ai harness, yet the two different models produced basically identical results, and questions the evaluation method.

The poster says Gemini 4 Pro's SVG output beats GPT-6 Astra and Claude Fable, while citing a LuminaBench post that the model's release has been delayed.

In an interview, Google DeepMind's Koray Kavukcuoglu says being at the frontier is the team's only goal and that he is 100% certain it will get there, with a link to a roundup of Gemini 4 progress.

The author posts a video captioned only "Gemini 4 Pro," without describing what was tested or any conclusion.

The author says Gemini 4 Pro is very strong in areas like design and 3D modeling but rough in others, because Google specializes models for fields like math rather than building all-round general models.

LuminaBench replies that Gemini 4 Pro can generate exploded views showing internal components with part names.

An unidentified Google model has surfaced in Arena testing, rumored to have 10M-token context, a big coding jump and persistent memory, possibly a Gemini 4 checkpoint.

A model reportedly routed through a gemini-3.7-flash Arena label generated a full interactive SVG product page with color variants, draggable controls and SVG download; Google has not confirmed its identity.