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

A user reports testing Gemini 4 Argon to build a 3D Gothic cathedral from a single image, using the post to seek mutual follows.

A user shares a new AI stack—Meta AI for chatting, Muse AI for agentic work, and Gemini 4 Argon for coding—after cancelling ChatGPT and Claude subscriptions.

A Japanese humor post personifies Gemini 4 Argon and jokes that it needs rescuing when it gets weak.

LuminaBench's 3D helicopter test: Carbon finished in Arena in just over 30 minutes, Opus 5.5 took 55 minutes in Claude Code.

A Japanese VoiSona creator says Gemini 4 Argon's vocal blend for the Amagai voice still falls slightly short, so manual tuning is needed before going fully Amagai.

LuminaBench generated a 3D helicopter with Gemini 4 Carbon in Arena in about 30 minutes, calling its 3D and engineering understanding clearly ahead of Kimi K3.1 Pro and Qwen 4, with an Opus 5.5 comparison coming.

The post says building on Gemini 4 Argon gets twice as expensive on Dec 1 at $4/$20 per million tokens, while Kimi K3.1 Pro goes open weights Nov 4.

A video compares Kimi K3.1, Opus 5.5, Gemini 4 Argon and GPT-6 Astra to see which performs better.

LisanBench added multi-reasoning-effort tests: Gemini 4 Argon low and medium emit no reasoning tokens yet score high, with low taking #1 on the non-thinking board despite being a dirty looper, and Argon high ranking second without the edit-distance rule.

HarshithLucky3 shows an early-stage attempt at driving Blender via MCP with Gemini 4 Argon, an initial foray into 3D modeling.

A tester says Kimi K3.1 Pro is a big quality jump with fewer hallucinations and lands around Gemini 4 Argon, with each model winning different tasks.

A user measured that the $100 Google AI Ultra plan maps to roughly $140 of Gemini 4 Argon usage, mocking the value versus Codex and Claude quotas and calling Gemini 3.8 Flash a freebie.