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

Two critical PHP flaws allowing remote server takeover without login were found with AI help including Gemini 4 Argon and GPT-5.6 Sol, prompting preventive site shutdowns worldwide.

A deep dive into two chained PHP CVEs enabling unauthenticated code execution notes that Wiz found the same flaw independently using GPT-5.6 Sol and Gemini 4 Argon.

A video compares the low-tier settings of Gemini 4 Carbon, Claude Opus 5.5 and GPT-6.1 Sol side by side.

Google CEO publicly confirms for the first time that internally Gemini 5 is already seen as far surpassing the just-launched Gemini 4 Argon.

The author asked Gemini 4 Argon for a Wispr Flow-like Mac voice app that talks to a Grok bot; try one was a web app, try two a working but ugly native build, and try three a Dieter Rams-inspired cleanup.

Artificial Analysis' personality benchmark finds Gemini models most confident and least likely to admit uncertainty, with Gemini 4 Argon at 97 confidence and 4 on uncertainty admission.

Google demos feeding LiDAR scans and archaeological archives into Antigravity, using Gemini 4 Argon and Blender MCP to build an explorable 3D heritage site.

A short video clip demonstrates what Gemini 4 Carbon produces on Arena, with the caption only noting it is running there.

A user had Gemini 4 Carbon generate a Cola SVG in Arena with a rough harness; the five-minute result was judged better than Opus.

Minimax M3.1 Flash, Astra 6, Argon 4, DeepSeek v4.1 Flash and Opus 5.5 each get 4 hours to improve a GTA 6-style Singapore demo, with 12-hour progress shown.

On the same prompt and effort level, Gemini 4 Argon finished the lighthouse scene in 1h47m for $6.30, far cheaper and faster than Sonnet 5.5, Opus 5.5 and GPT-6 Astra.

A tester says Gemini 4 Carbon excels at 3D but lags on SVG output, taking longer than Gemini 4 Argon while delivering better quality.