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

A roundup of rumors about Google's internal Gemini Carbon checkpoint: possibly an Argon upgrade with big SWE-bench and DeepSWE gains over Barium-B, compared by an employee to Opus 5.5, in wider internal A/B tests and possibly on Arena, but 2–4x slower than Argon.

A Japanese post outlines five steps for using Gemini 4 Argon to build a custom work tool that automates routine checks and tallying.

While inspecting Antigravity's backend via CC-switch, a user found requests routed to gemini-3.8-flash-tiered and gemini-3.8-flash-n instead of a fixed model, guessing tiered is a canary routing model for new releases.

A user prompted Gemini 4 Argon to build a Demon Slayer Infinity Castle 3D scene, shared the video, and judged Argon possibly better than Carbon at this task.

Security firm Calif used GPT-5.6 Sol and Wiz used Gemini 4 Argon to independently find and demonstrate a PHP 5.4+ memory-reuse vulnerability, with AI automating the memory-layout tuning that was previously manual.

Short on AI quota, the user added a Claude account and now runs Claude 4, Codex 2 and Antigravity (Gemini, Argon), saying Gemini and Argon quietly do a lot of the work and plans a webinar on splitting tasks between them.

A user jokes that Gemini 4 Argon drew a detailed takedown, a T3 PR ban, and a 'worst release of the year' label.

A user says Gemini 4 Argon burns through the Pro plan's 5-hour limit after a few ordinary chat messages, calling it unusable for coding or agentic work.

A roundup says Carbon codes near Claude Opus 5.5, beats Argon on 3D and spatial reasoning, is on LM Arena, and may ship late October or November.

Using a $9 three-month Gemini Pro plan, the author had Gemini 4 Argon in Antigravity render a photoreal fish tank image into 3D, finding it far better than the web version, and shared the prompt.

The author used Gemini 4 Argon to build a working interactive 3D Indonesian neighborhood simulator with NPCs, traffic, random incidents and ghost events; visuals and UI still need work.

The poster ran a US-vs-Iran chess demo on a Gemini 4 Carbon Arena checkpoint (id: 01a161e8…) and says it clearly outperforms Kimi K3.1 Pro, with video attached.