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

Google DeepMind used Gemini Argon to reduce Turing machine state counts for Goldbach's Conjecture, the Riemann Hypothesis and ZF consistency, verifying each reduction end-to-end in Lean 4 and open-sourcing the code.

The poster expects Google DeepMind's Gemini 4 Carbon soon and debates whether it ships as Gemini 4.1 Carbon or a new class in the 4.0 family.

Citing Toby's finding, Gemini 4 (Google Argon) sustains real-world AI research tasks for about 15 minutes, far below the 6-hour METR and 18-hour AI 2027 forecasts.

A user asks Gemini 4 Carbon to build a 3D bicycle riding a cow in three.js, praising the 3D output and comparing it to Astra and Opus 5.5.

The author had Gemini 4 Argon build a periodic table synced to Tom Lehrer's 'The Elements', then gave the same prompt to Opus 5.5 and posted a video comparing the two results.

A developer implemented the robotics harness described in Gemini 4 Argon's report; Gemini completed a zero-shot bimanual handover, and the code is now open source with a demo video.

The poster argues OpenAI is only shipping point updates like gpt-6.1 sol, while Google has shipped Argon and is reportedly already working on Carbon.

The author tested a Gemini 4 Carbon checkpoint labeled gemini-4-argon in Arena, generating an interactive 3D crystal citrus in about 20 minutes and judging its 3D capability better than Argon.

Google releases Gemini 4 Argon worldwide, reporting over 35% more agent-based code submissions via Antigravity and 87% on the ERQA embodied reasoning test.

Pichai says Gemini 4 Argon burned nearly 25 trillion tokens in its first 72 hours, 3x Gemini 3.0, while Google rents ~110K Nvidia GPUs from SpaceX for $920M a month.

A user posts the classic pelican-riding-a-bicycle SVG generated by Gemini 4 Carbon, another Gemini 4 model name alongside Argon.

Greg Speicher calls a take that judges Google by Argon superficial and biased, citing Search, YouTube, Cloud, Waymo and TPUs.